SlideShare a Scribd company logo

Hardware and Software Architectures for the CELL BROADBAND ENGINE processor

S
Slide_N

This document provides an overview of the Cell Broadband Engine processor architecture. It describes the key components of the Cell processor including the Power Processor Element (PPE) and 8 Synergistic Processor Elements (SPEs). The PPE acts as a general purpose 64-bit RISC processor, while the SPEs are intended for intensive numeric computation. The document outlines the memory hierarchy and DMA capabilities that provide high bandwidth memory access. It also describes the internal Element Interconnect Bus and external interfaces that enable high bandwidth communication on the chip.

Technology
Related topics:
Cell Broadband Engine
1 of 87
Downloaded 16 times
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
IBM Systems and Technology Group © 2005 IBM Corporation Tutorial Hardware and Software Architectures for the CELL BROADBAND ENGINE processor Michael Day, Peter Hofstee IBM Systems & Technology Group, Austin, Texas CODES+ISSS Conference, September 2005
2 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Agenda Trends in Processors and Systems Introduction to Cell Challenges to processor performance Cell Broadband Engine Architecture (CBEA) Cell Broadband Engine (CBE) Processor Overview Cell Programming Models Prototype Software Environment TRE Demo © 2005 IBM Corporation
3 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Trends in Microprocessors and Systems
4 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Processor Performance over Time (Game processors take the lead on media performance) 10 100 1000 10000 100000 1000000 1993 1995 1996 1998 2000 2002 2004 Game Processors PC Processors Flops (SP)
5 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation System Trends toward Integration Implied loss of system configuration flexibility Must compensate with generality of acceleration function to maintain market size. Processor Southbridge NorthbridgeMemory Accel Next-Gen Processor Memory IO IO
6 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Motivation: Cell Goals Outstanding performance, especially on game/multimedia applications. • Challenges: Memory Wall, Power Wall, Frequency Wall Real time responsiveness to the user and the network. • Challenges: Real-time in an SMP environment, Security Applicable to a wide range of platforms. • Challenge: Maintain programmability while increasing performance Support an introduction in 2005/6. • Challenge: Structure innovation such that 5yr. schedule can be met
7 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Performance Limiters in Conventional Microprocessors Memory Wall • Latency induced bandwidth limitations Power Wall • Must improve efficiency and performance equally Frequency Wall • Diminishing returns from deeper pipelines (can be negative if power is taken into account)
8 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell © 2005 IBM Corporation Systems and Technology Group
9 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell History IBM, SCEI/Sony, Toshiba Alliance formed in 2000 Design Center opened in March 2001 Based in Austin, Texas February 7, 2005: First external technical disclosures • Cell Broadband Engine Architecture documentation can be found at: http://www.ibm.com/developerworks/power/cell • Additional publications on Cell can be downloaded from: http://www.ibm.com/chips/techlib/techlib.nsf/products/Cell • A paper on Cell in the upcoming issue of the IBM Journal of Research and Development can be found at: http://www.research.ibm.com/journal/rd/494/kahle.html Systems and Technology Group © 2005 IBM Corporation
10 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Highlights Supercomputer on a chip Multi-core microprocessor (9 cores) 3.2 GHz clock frequency 10x performance for many applications Digital home to distributed computing Systems and Technology Group © 2005 IBM Corporation
11 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Introducing Cell Sets a new performance standard • Exploits parallelism while achieving high frequency • Supercomputer attributes with extreme floating point capabilities • Sustains high memory bandwidth with smart DMA controllers Designed for natural human interaction • Photo-realistic effects • Predictable real-time response • Virtualized resources for concurrent activities Designed for flexibility • Wide variety of application domains • Highly abstracted to highly exploitable programming models • Reconfigurable I/O interfaces • Autonomic power management
12 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Microprocessor Architecture Trends CPU GPU NIC Security Media … CPU Streaming Graphics Processor Network Processor Security Processor Media Processor … 64b Power Processor Synergistic Processor Synergistic Processor ... Mem. Contr. Config. IO Hardwired Function Programmable ASIC Cell
13 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Key Attributes of Cell Cell is Multi-Core • Contains 64-bit Power Architecture TM • Contains 8 Synergistic Processor Elements (SPE) Cell is a Flexible Architecture • Multi-OS support (including Linux) with Virtualization technology • Path for OS, legacy apps, and software development Cell is a Broadband Architecture • SPE is RISC architecture with SIMD organization and Local Store • 128+ concurrent transactions to memory per processor Cell is a Real-Time Architecture • Resource allocation (for Bandwidth Measurement) • Locking Caches (via Replacement Management Tables) Cell is a Security Enabled Architecture • SPE dynamically reconfigurable as secure processors
14 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Architecture is … COHERENT BUS Power ISA MMU/BIU Power ISA MMU/BIU … IO transl. Memory Incl. coherence/memory compatible with 32/64b Power Arch. Applications and OS’s 64b Power Architecture™
15 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Architecture is … 64b Power Architecture™ COHERENT BUS (+RAG) Power ISA +RMT MMU/BIU +RMT Power ISA +RMT MMU/BIU +RMT IO transl. Memory Plus Memory Flow Control (MFC) MMU/DMA +RMT Local Store Memory MMU/DMA +RMT Local Store Memory LS Alias LS Alias … … …
16 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Architecture is … 64b Power Architecture™+ MFC COHERENT BUS (+RAG) Power ISA +RMT MMU/BIU +RMT Power ISA +RMT MMU/BIU +RMT IO transl. Memory Plus Synergistic Processors MMU/DMA +RMT Local Store Memory MMU/DMA +RMT Local Store Memory LS Alias LS Alias … … … Syn. Proc. ISA Syn. Proc. ISA
17 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell - Attacking the Performance Walls Multi-Core Non-Homogeneous Architecture • Control Plane vs. Data Plane processors • Attacks Power Wall 3-level Model of Memory • Main Memory, Local Store, Registers • Attacks Memory Wall Large Shared Register File & SW Controlled Branching • Allows deeper pipelines • Attacks Frequency Wall Systems and Technology Group © 2005 IBM Corporation
18 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Frequency Increase vs Power Consumption 0 0.5 1 1.5 2 2.5 3 3.5 0.9 1 1.1 1.2 1.3 Voltage Realative Power Frequency
19 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Power Efficient Architecture and the CellBE Non-Homogeneous Coherent Multi-Processor • Data-plane/Control-plane specialization • More efficient than homogeneous SMP 3-level model of Memory • Bandwidth without (inefficient) speculation • High-bandwidth .. Low power
20 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Power Efficient Architecture and the SPE Power Efficient ISA allows Simple Control • Single mode architecture • No cache • Branch hint • Large unified register file • Channel Interface Efficient Microarchitecture • Single port local store • Extensive clock gating Efficient implementation • See Cool Chips paper by O. Takahashi et al. and T. Asano et al.
21 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Broadband Engine Components
22 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Processor Components Power Core (PPE) L2 Cache NCU Power Processor Element (PPE): •General Purpose, 64-bit RISC Processor (PowerPC AS 2.0.2) •2-Way Hardware Multithreaded •L1 : 32KB I ; 32KB D •L2 : 512KB •Coherent load/store •VMX •3+ GHz •Real-time Control •Locking L2 Cache & TLB •Bandwidth Reservation In the Beginning – the solitary Power Processor Custom Designed – for high frequency, space and power efficiency
23 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation 96 Byte/Cycle 300+GB/sec @ 3.2 GHz Element Interconnect Bus Cell Processor Components Power Core (PPE) L2 Cache NCU N N N MIC EIB data ring for internal communication • Four 16 byte data rings, supporting multiple transfers • 96B/cycle peak bandwidth • Over 100 outstanding requests • 300+ GByte/sec @ 3.2 GHz
24 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation 96 Byte/Cycle 300+ GB/sec @ 3.2 GHz Element Interconnect Bus Cell Processor Components Power Core (PPE) L2 Cache NCU N N N MIC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC Local Store SPUMFC N AUC Local Store SPUMFC N AUC SPE provides computational performance • Dual issue, up to 8-way 32-bit SIMD • Dedicated resources: 128 128-bit RF, 256KB Local Store • Each DMA MMU can be dynamically configured to protect resources • Dedicated DMA engine: Up to 16 outstanding requests Local Store SPU MFC N AUC Local Store SPU MFC N AUC
25 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation 96 Byte/Cycle 300+ GB/sec @ 3.2GHz Element Interconnect Bus Cell Processor Components Power Core (PPE) L2 Cache NCU N N N MIC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC Local Store SPUMFC N AUC Local Store SPUMFC N AUC Memory Flow Controller (MFC) and Atomic Update Cache (AUC) • 4 line cache for shared memory atomic update primitives •High performance DMA unit •Local Store aliased into PPE system memory •16 element SPE side DMA Command Queue • 8 element PPE side DMA Command Queue •DMA List supports 2K entry scatter/gather •MFC-MMU controls SPE DMA accesses •Compatible with PowerPC Virtual Memory architecture •Full memory protection for MFC DMA •S/W controllable from PPE MMIO •DMA 1,2,4,8,16,128 -> 16Kbyte transfers for I/O access
26 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Processor Components IOIF0 20 GB/sec BIF or IOIF0 IOIF1 Southbridge I/O 5 GB/sec LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC Local Store SPUMFC N AUC Local Store SPUMFC N AUC Local Store SPU MFC N AUC Local Store SPU MFC N AUC 96 Byte/Cycle Element Interconnect Bus Power Core (PPE) L2 Cache NCU Memory Interface Controller (MIC): • Dual XDRTM controller (25.6GB/s @ 3.2Gbps) • ECC support • Suspend to DRAM support Broadband Interface Controller (BIC): Provides a wide connection to external devices Two configurable interfaces (50+GB/s @ 5Gbps) – Configurable number of bytes – Coherent (BIF) and / or I/O (IOIFx) protocols Supports two virtual channels per interface Supports multiple system configurations MIC 25 GB/sec XDR DRAM
27 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Processor Components I/O Bus Master Translation (IOT) Translates Bus Addresses to System Real Addresses Two Level Translation – I/O Segments (256 MB) – I/O Pages (4K, 64K, 1M, 16M byte) I/O Device Identifier per page for LPAR IOST and IOPT Cache – hardware / software managed IOIF0 20 GB/sec BIF or IOIF0 MIC 25 GB/sec XDR DRAM IOIF1 Southbridge I/O 5 GB/sec LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC Local Store SPUMFC N AUC Local Store SPUMFC N AUC Local Store SPU MFC N AUC Local Store SPU MFC N AUC 96 Byte/Cycle Element Interconnect Bus Power Core (PPE) L2 Cache NCU IIC IOT Internal Interrupt Controller (IIC) Handles SPE Interrupts Handles External Interrupts – From Coherent Interconnect – From IOIF0 or IOIF1 Interrupt Priority Level Control Interrupt Generation Ports for IPI Duplicated for each PPE hardware thread
28 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation 96 Byte/Cycle 256GB/sec @ 4GHz Element Interconnect Bus Cell Processor Components Power Core (PPE) L2 Cache NCU N N LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC Local Store SPUMFC N AUC Local Store SPUMFC N AUC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC MIC IOIF1 IOIF0 Southbridge I/O 20 GB / sec BIF or IOIF1 25 GB / sec XDR DRAM 5 GB/sec Token Manager – Bandwidth Reservation for shared resources • Optionally Enabled for RT Tasks or LPAR • Multiple Resource Allocation Groups • Generates access tokens at configurable rate for each allocation group 1 per each memory bank (16 total) 2 for each IOIF (4 total) • Requestors assigned RAG ID by OS/hypervisor Each SPE PPE L2 / NCU IOIF 0 Bus Master IOIF 1 Bus Master • Priority order for using another RAGs unused tokens • Resource overcommit warning interrupt IIC IOTTKM
29 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell ProcessorCell Processor Power Processor Element (PPE): •General Purpose, 64-bit RISC Processor (PowerPC AS 2.0) •2-Way Hardware Multithreaded •L1 : 32KB I ; 32KB D •L2 : 512KB •Coherent load/store •VMX •3.2 GHz Internal Interconnect: •Coherent ring structure •300+ GB/s total internal interconnect bandwidth •DMA control to/from SPEs supports >100 outstanding memory requests Synergistic Processor Elements (SPE): •8 per chip •128-bit wide SIMD Units •Integer and Floating Point capable •256KB Local Store •Up to 25.6 GF/s per SPE --- 200GF/s total * External Interconnects: •25.6 GB/sec BW memory interface •2 Configurable I/O Interfaces •Coherent interface (SMP) •Normal I/O interface (I/O & Graphics) •Total BW configurable between interfaces •Up to 35 GB/s out •Up to 25 GB/s in Memory Management & Mapping •SPE Local Store aliased into PPE system memory •MFC/MMU controls SPE DMA accesses •Compatible with PowerPC Virtual Memory architecture •S/W controllable from PPE MMIO •Hardware or Software TLB management •SPE DMA access protected by MFC/MMU ““Supercomputer-on-a-Chip” * At clock speed of 3.2GHz Element Interconnect Bus MFC Local Store SPU N AUC MFC Local Store SPU N AUC Power Core (PPE) L2 Cache NCU Local Store SPU MFC N AUC Local Store SPU MFC N AUC N N N N MFC LocalStore SPU N AUC MFC LocalStore SPU N AUC MFC LocalStore SPU N AUC MFC LocalStore SPU N AUC 20 GB/sec Coherent Interconnect 25.6 GB/sec Memory Inteface 5 GB/sec I/O Bus
30 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation SPE Highlights RISC like organization • 32 bit fixed instructions • Clean design – unified Register file User-mode architecture • No translation/protection within SPU • DMA is full Power Arch protect/x-late VMX-like SIMD dataflow • Broad set of operations (8 / 16 / 32 Byte) • Graphics SP-Float • IEEE DP-Float Unified register file • 128 entry x 128 bit 256KB Local Store • Combined I & D • 16B/cycle L/S bandwidth • 128B/cycle DMA bandwidth LS LS LS LS GPR FXU ODD FXU EVN SFP DP CONTROL CHANNEL DMA SMM ATO SBI RTB FWD 14.5mm2 (90nm SOI)
31 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation What is a Synergistic Processor? (and why is it efficient?) Local Store “is” large 2nd level register file / private instruction store instead of cache • Asynchronous transfer (DMA) to shared memory • Frontal attack on the Memory Wall Media Unit turned into a Processor • Unified (large) Register File • 128 entry x 128 bit Media & Compute optimized • One context • SIMD architecture LS LS LS LS GPR FXU ODD FXU EVN SFP DP CONTROL CHANNEL DMA SMM ATO SBI RTB FWD SPU SMF
32 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation SPE Detail LS LS LS LS GPR FXU ODD FXU EVN SFP DP CONTROL CHANNEL DMA SMM ATO SBI RTB FWD Synergistic Processor Element (SPE) User-mode architecture • No translation/protection within SPE • DMA is full PowerPC protect/xlate Direct programmer control • DMA/DMA-list • Branch hint VMX-like SIMD dataflow • Graphics SP-Float • No saturate arith, some byte • IEEE DP-Float (BlueGene- like) Unified register file • 128 entry x 128 bit 256KB Local Store • Combined I & D • 16B/cycle L/S bandwidth • 128B/cycle DMA bandwidth SPE Latencies • Simple fixed point - 2 cycles* • Complex fixed point - 4 cycles* • Load - 6 cycles* Local store size = 256 KB • Single-precision (ER) float - 6 cycles* • Integer multiply - 7 cycles* • Branch miss - 20 cycles No penalty if correctly hinted • DP (IEEE) float - 13 cycles* Partially pipelined • Enqueue DMA Command - 20 cycles* SPU Units: • Simple (FXU even) Add/Compare Rotate Logical, Count Leading Zero • Permute (FXU odd) Permute Table-lookup • FPU (Single / Double Precision) • Control (SCN) Dual Issue, Load/Store, ECC Handling • Channel (SSC) – Interface to MFC • Register File (GPR/FWD)
33 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Coherent Offload Model DMA into and out of Local Store similar to Power core loads & stores Governed by Power Architecture page and segment tables for translation and protection Shared memory model • Power architecture compatible addressing • MMIO capabilities for SPEs • Local Store is mapped (alias) allowing LS to LS DMA transfers • DMA equivalents of locking loads & stores • OS management/virtualization of SPEs Pre-emptive context switch is supported
34 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation MFC Detail Memory Flow Control System •DMA Unit •LS <-> LS, LS<-> Sys Memory, LS<-> I/O Transfers •8 PPE-side Command Queue entries •16 SPU-side Command Queue entries •MMU similar to PowerPC MMU •8 SLBs, 256 TLBs •multiple page sizes •Software/HW page table walk •PT/SLB misses interrupt PPE •Atomic Cache Facility •4 cache lines for atomic updates •2 cache lines for cast out/MMU reload •Up to 16 outstanding DMA requests in BIU •Resource / Bandwidth Management Tables •Token Based Bus Access Management •TLB Locking Isolation Mode Support (Security Feature) Hardware enforced “isolation” • SPE and Local Store not visible (bus or jtag) • Small LS “untrusted area” for communication area Secure Boot • Chip Specific Key • Decrypt/Authenticate Boot code “Secure Vault” – Runtime Isolation Support • Isolate Load Feature • Isolate Exit Feature Legend: Data Bus Snoop Bus Control Bus Xlate Ld/St MMIO Local Store SPU DMA Engine DMA Queue Atomic Facility MMU RMT Bus I/F Control SPE MMIO
35 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Implementation Aspects
36 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Pre-Decode L1 Instruction Cache Microcode SMT Dispatch (Queue) Decode Dependency Issue Branch Scan Fetch Control L2 Interface VMX/FPU Issue (Queue) VMX Load/Store/ Permute VMX Arith./Logic Unit FPU Load/Store FPU Arith/Logic Unit Load/Store Unit Branch Execution Unit Fixed-Point Unit FPU CompletionVMX Completion Completion/Flush 8 4 2 1 2 11 1 1 111 4 Thread A Thread B Threads alternate fetch and dispatch cycles Thread A Thread B Thread A L1 Data Cache 2 PPE BLOCK DIAGRAM
37 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation ID2 IS1 RF2 EX1 EX2 IC1 IC2 IC3 ID1 RF1 RF2 EX1 EX3 EX4EX2 MC1 IS2 RF1 RF2 EX1 EX2 EX3 EX4 Fixed Point Unit Instruction PPE PIPELINE FRONT END PPE PIPELINE BACK END Microcode MC2 MC3 MC4 EX5 EX5 ... MC9 MC10 MC11 IB2 IS3 EX7 WB EX6 IB1 ID3 EX3RF1DLY DLY DLY DLY DLY DLY EX4 IBZ IC0 Branch Instruction EX8 WB IC4 BP1 BP2 BP3 BP4 Load/Store Instruction Branch Prediction Instruction Decode and IssueInstruction Cache and Buffer IC Instruction Cache IB Instruction Buffer BP Branch Prediction MC Microcode ID Instruction Decode IS Instruction Issue DLY Delay Stage RF Register File Access EX Execution WB Write Back
38 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Permute Unit Load-Store Unit Floating-Point Unit Fixed-Point Unit Branch Unit Channel Unit Result Forwarding and Staging Register File Local Store (256kB) Single Port SRAM 128B Read 128B Write DMA Unit Instruction Issue Unit / Instruction Line Buffer 8 Byte/Cycle 16 Byte/Cycle 128 Byte/Cycle64 Byte/Cycle On-Chip Coherent Bus SPE BLOCK DIAGRAM
39 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation EX1 EX3 EX4EX2 SPU PIPELINE FRONT END SPE PIPELINE BACK END EX5 EX6 RF1 RF2 Branch Instruction WB Load/Store Instruction IF Instruction Fetch IB Instruction Buffer ID Instruction Decode IS Instruction Issue RF Register File Access EX Execution WB Write Back IF1 IF2 ID2 IS1IF3 IF4 IF5 ID1 IS2IB2IB1 ID3 EX2 Fixed Point Instruction WBEX1 Floating Point Instruction WBEX1 EX2 Permute Instruction WBEX1 EX3 EX4 EX5 EX6EX2 EX3 EX4
40 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation CellBE Processor ~250M transistors ~235mm2 Top frequency >3GHz 9 cores, 10 threads > 200+ GFlops (SP) @3.2 GHz > 20+ GFlops (DP) @3.2 GHz Up to 25.6GB/s memory B/W Up to 50+ GB/s I/O B/W ~400M$(US) design investment
41 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Processor Can Support Many Systems Game console systems Blades HDTV Home media servers Supercomputers Cell BE Processor XDRtm XDRtm IOIF0 IOIF1 Cell BE Processor XDRtm XDRtm IOIF BIF Cell BE Processor XDRtm XDRtm IOIF Cell BE Processor XDRtm XDRtm IOIF BIF Cell BE Processor XDRtm XDRtm IOIF CellBE Processor XDRtm XDRtm IOIF BIF CellBE Processor XDRtm XDRtm IOIFSW
42 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell BE Processor Initial Application Areas Cell excels at processing of rich media content in the context of broad connectivity • Digital content creation (games and movies) • Game playing and game serving • Distribution of dynamic, media rich content • Imaging and image processing • Image analysis (e.g. video surveillance) • Next-generation physics-based visualization • Video conferencing • Streaming applications (codecs etc.) • Physical simulation & science Cell is an excellent match for any applications that require: • Parallel processing • Real time processing • Graphics content creation or rendering • Pattern matching • High-performance SIMD capabilities
43 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Programmability Programmer Productivity Raw Hardware Performance BE Design 25 GB/sec Memory Inteface 20 GB/sec Coherent Interconnect 5 GB/sec I/O Bus Cell Software Overview 256 GB/sec Coherent Ring MFC Local Store SPU N AUC MFC Local Store SPU N AUC Power Processor (PPE) L2 Cache NCU Local Store SPU MFC N AUC Local Store SPU MFC N AUC N N N N Cell Chip MFC LocalStore SPU N AUC MFC LocalStore SPU N AUC MFC LocalStore SPU N AUC MFC LocalStore SPU N AUC Cell Programming Features Flexible Programming Models Function Offload Model Existing Application Accelerator Parallel Computational Acceleration Heterogeneous Multi-Threading Runtime
44 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Programming Characteristics Exploit all of Cell’s 18 asynchronous engines • Through function offload, or parallel computational tasks Decompose work into data parallel blocks • Independent Data parallel tasks are most efficient Overlap SPE compute with DMA loads and stores • Multibuffering , pipelining Reduce PPE/SPE workload ratio • PPE as Control / OS processor • SPE as heavy lifting computational engines Super-Linear speedups over conventional processors may be achieved • Through exploitation of asynchronous DMA engines
45 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Features Exploited by Software Keeping Intermediate/Control Data on-Chip MMU – ERATS, SLBs, TLBs DMA from L2 cache-> LS LS to LS DMA Cache <-> Cache transfers (atomic update) SPE Signalling Registers SPE <-> PPE Mailboxes Resource Reservation and Allocation PPE can be shared across logical partitions SPEs can be assigned to logical partitions SPEs independently or Group Allocated Cache Replacement Management TLB Replacement Management Bandwidth Reservation Cell Programming Features Power Processor (PPE) L2 Cache MFC Local Store SPU N AUC DMA with Intervention Atomic Update Cache Local Store to Local Store DMA Hardware Managed Cache Coherency Cell Chip System Memory I/O MFC Local Store SPU N AUC MFC Local Store SPU N AUC MFC Local Store SPU N AUC MFC Local Store SPU N AUC MFC Local Store SPU N AUC BIF/IOIF
46 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Subsystem Programming Model Dedicated Function (problem/privileged subsystem) Programmer writes/uses SPU "libraries" Graphics Pipeline Audio Processing MPEG Encoding/Decoding Encryption / Decryption Main Application in PPE, invokes SPU bound services RPC Like Function Call I/O Device Like Interface (FIFO/ Command Queue) 1 or more SPUs cooperating in subsystem Problem State (Application Allocated) –Transcoding –Realtime data transformation & streaming –Graphics Processing –MPEG Encoding / Decoding –Physical Simulation Privileged State (OS Allocated) –Encryption / Decryption Services –Network Packet Filtering / Routing Code-to-data or data-to-code pipelining possible Very efficient in real-time data streaming applications Function Offload Power Core (PPE) Sys. Memory/QoS SPU Local Store MFC N SPU Local Store MFC N Multi-stage Pipeline SPU Local Store MFC N SPU Local Store MFC N SPU Local Store MFC N Parallel-stages Power Core (PPE) Sys. Memory/QoS
47 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Application Specific Acceleration ModelApplication Specific Acceleration Model –– SPC Accelerated SubsystemsSPC Accelerated Subsystems System Memory Parameter Area PPE PowerPC Application Cell Aware OS (Linux) Compression/ Decompression Realtime MPEG Encoding TCP Offload Data Encryption Data Decryption Conventional PowerPC App. with Library Dependencies OpenGL Encrypt Decrypt Encoding OS loads 3rd party function libraries containing PPC wrapped SPU code PPE part of library code loads SPU allocated by OS Loads SPU part of library code PPC part of library App calls library function, PPE library stub communications to SPEmpeg_encode() MFC Local Store SPU N AUC MFC Local Store SPU N AUC MFC Local Store SPU N AUC MFC Local Store SPU N AUC MFC Local Store SPU N AUC MFC Local Store SPU N AUC MFC Local Store SPU N AUC MFC Local Store SPU N AUC Maintains PowerPC programming model Similiar to IOP but integrated with main processor • Supports shared memory programming • Extremely high on-chip bandwidth • Scales with “conventional” processor SPU code provided in: • Application or middleware libraries • Operating System Services Does not require application rewrite • Specific subsystems targetted • Separate compilations • Leverage ELF • SPUs managed by OS SPE Programming localized in library • SPU Code vectorization required • Private Local Store and DMA model •Scheduling of code / data movement • Code debug and fault isolation complexity Application Specific Accelerators Decoding
48 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Parallel Computational Acceleration Single Source Compiler (PPE and SPE targets) • Auto parallelization ( treat target Cell as an Shared Memory MP ) • Auto SIMD-ization ( SIMD-vectorization ) for PPE VMX and SPE • Compiler management of Local Store as Software managed cache (I&D) Optimization Options • OpenMP-like pragmas • MPI based Microtasking • Streaming languages • Vector.org SIMD intrinsics • Data/Code partitioning • Streaming / pre-specifying code/data use Compiler or Programmer scheduling of DMAs Compiler use of Local store as soft-cache IBM Research Prototype Single Source Compiler • C Frontend • XLC SPE and XLC PPE back-end IBM Research Prototype Parallelizing Compiler • UPC front-end • Fortran front-end • XLC SPE backend 256 GB/sec Coherent Ring MFC Local Store SPU N AU C MFC Local Store SPU N AU C Power Processor (PPE) L2 Cache NCU Local Store SPU MFC N AU C Local Store SPU MFC N AU C N N N N Broadband Engine Chip MFC Local Store SPU N AU C MFC Local Store SPU N AU C MFC Local Store SPU N AU C MFC Local Store SPU N AU C 256 GB/sec Coherent Ring MFC Local Store SPU N AU C MFC Local Store SPU N AU C Power Processor (PPE) L2 Cache NCU Local Store SPU MFC N AU C Local Store SPU MFC N AU C N N N N Broadband Engine Chip MFC Local Store SPU N AU C MFC Local Store SPU N AU C MFC Local Store SPU N AU C MFC Local Store SPU N AU C XDRtm XDRtm I/O I/O Tools and Programmer Driven
49 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Application Source & Libraries PPE object files SPE object files SPESPE SPE SPE Physical SPEs SPESPE SPE SPE Cell AwareOS ( Linux) SPE Virtualization / Scheduling Layer (m->n spe lwp/threads) New SPE lwp/threadsExisting PPE lwp/threads Physical PPE PPE MT1 MT2 Operating System Runtime Strategy Heterogeneous Multi-Threading Model PPE LWP/Threads SPU LWP/Threads SPU DMA EA = PPE Process EA Space OS supports Create/Destroy SPE tasks Atomic Update Primitives used for Mutex SPE Context Fully Managed Context Save/Restore for Debug Virtualization Mode (indirect access) Direct Access Mode (realtime) OS assignment of SPE threads to SPUs Programmer directed using affinity mask
50 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Prototype Cell Extensions to Linux System Call Interface exec Loader File System Framework Device Framework Network Framework Streams Framework SPU Management Framework Privileged Kernel Extensions Firmware / Hypervisor ILP32 Processes LP64 EM Processes Cell Reference System Hardware 32-bit GNU Libs (glibc,etc) 64-bit Linux Kernel 64-bit GNU Libs (glibc) SPUFS FilesystemMisc format bin SPU Object Loader Extension ptrace, large page, madvise, SPE error, fault handling, IIC support, IOS Architecture Specific Code SPU Allocation, Scheduling & Dispatch Extension PPC32 Apps. Cell32 Workloads PPC64 Apps.Cell64 Workloads std. PPC32 elf interp SPE Object Loader Services SPE Management Runtime Library (32-bit) std. PPC64 elf interp Programming Models Offered: RPC, Device Subsystem, Non-Realtime and Realtime Asymmetric Threads -- Single SPU, SPU Groups, Shared Memory SPE Management Runtime Library (64-bit)
51 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Programmer Experience Development Tools Stack System Level Simulator Linux PPC64 with Cell Extensions SPE Management Lib Application Libs Samples Workloads Demos Code Dev Tools Debug Tools Performance Tools Standards: Language extensions ABI Verification Hypervisor Development Environment End-User Experience Execution Environment Cell Prototype Software Environment Miscellaneous Tools
52 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Standards Application Binary Interface Specifications • Defines such things as data types, register usage, calling conventions, and object formats to ensure compatibility of code generators and portability of code. SPE ABI Linux Cell ABI SPE C/C++ Language Extensions • Defines standardized data types, compiler directives, and language intrinsics used to exploit SIMD capabilities in the core. • Data types and Intrinsics styled to be similar to Altivec/VMX. SPE Assembly Language Specification Standards
53 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation System Level Simulator Cell BE – full system simulator • Uni-Cell and multi-Cell simulation • User Interfaces – TCL and GUI • Cycle accurate SPU simulation (pipeline mode) • Pseudo accurate memory and MFC modes • Emitter facility for tracing and viewing simulation events Other simulators • spusim – standalone SPU simulator Execution Environment
54 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Verification Hypervisor (vHype) Seamless Integration of Dual Environments •Low overhead – small footprint •Realtime Resource management •SPE management •Separate policy manager •Pre-emptive partition switching on high priority interrupts Real time Open Source / Proprietary OS Conventional Network OS (Linux) I/O Hosting Hypervisor (Hypervisor - Multi-Partition Resource Allocation / Reservation /Protection) Hardware Networked Apps. Inter-partition Communications Logical Partitioning RTOS/Linux Execution Environment
55 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Linux on Cell All software in STIDC written on Linux OS • Started with Linux 2.4 PPC64 on Cell Simulator SPEs exposed as I/O Devices (function offload model) SPE DMA required pre-pinned memory Inflexible programming model Moved to 2.6.3 • Added heterogenous lwp/thread model – via system call – moved to SPUFS in 2.6.13 SPE thread API created (similar to pthreads library) User mode direct and indirect SPE access models Full pre-emptive SPE context management spe_ptrace() added for gdb support spe_schedule() for thread to physical SPE assignment currently FIFO – run to completion • SPE threads share address space with parent PPE process (through DMA) Demand paging for SPE accesses Shared hardware page table with PPE • SPE Error, Event and Signal handling directed to parent PPE thread • SPE elf objects wrapped into PPE shared objects with extended gld SPE-side mini-loader • madvise() extended for L2 cache and TLB locking/preloading (realtime feature) • All patches for Cell in architecture dependent layer (subtree of PPC64) Publishing Initial CellBE Patches for 2.6.13 (Fall 2005 target) Execution Environment
56 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation SPE Management Library SPEs are exposed as threads • SPE thread model interface is similar to POSIX threads. • SPE thread consists of the local store, register file, program counter, and MFC-DMA queue. • Associated with a single Linux task. • Features include: Threads - create, groups, wait, kill, set affinity, set context Thread Queries - get local store pointer, get problem state pointer, get affinity, get context Groups - create, set group defaults, destroy, memory map/unmap, madvise. Group Queries - get priority, get policy, get threads, get max threads per group, get events. SPE image files - opening and closing SPE Executable • Standalone SPE program managed by a PPE executive. • Executive responsible for loading and executing SPE program. It also services “syscall” requests for I/O (eg, fopen, fwrite, fprintf) and memory requests (eg, mmap, shmat, …). Execution Environment
57 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Optimized Prototype Libraries Audio (resampling) Cryptographic Fast fourier transform Game math Image Large matrix Math Matrix (4x4) Miscellaneous Memory management Multi-precision math Noise & Turbulence Oscillator Parallel programming (MPI like) Shared memory SPU plugin SPU system call Surfaces / curves Synchronization Vector Execution Environment
58 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Code Development Tools GNU based binutils • gas SPE assembler • gld SPE ELF object linker gld extensions for embedding SPE object modules in PPE executables • misc bin utils (ar, nm, ...) targeting SPE modules • hosted on Linux IA32, Linux PowerPC GNU based C/C++ compiler targeting SPE • From STI Partner • retargeted compiler to SPE • Supports common SPE Language Extensions and ABI (ELF/Dwarf2) object output Cell Broadband Engine Optimizing Compiler (IBM Proprietary) • IBM XLC C/C++ for PowerPC (Tobey) • IBM XLC C retargeted to SPE assembler (including vector intrinsics) - highly optimizing • Prototype - XLC Compiler supporting CellBE Programmer Productivity Aids Single Source compilation using OpenMP like pragmas (PPE and SPE object code generated) Auto-Vectorization (auto-SIMD) for SPE code Auto-Parallelization across SPEs UPC Front end – parallelization across SPEs Local Store software managed caching model • Hosted on Linux • Executables to be available on IBM Alphaworks – Fall 2005 Development Environment
59 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Debug Tools CellBE system simulator • Executable availability on AlphaWorks (Fall 2005 target) GNU gdb • ptrace and spe_ptrace enabled • Multi-core Application source level debugger supporting PPE multithreading, SPE multithreading, interacting PPE and SPE threads • Three modes of debugging SPU threads Attach to SPE thread Launch mode – launch a new debug session for each SPE thread Pass-thru mode – follow execution into SPE thread RISCwatch • Low level hardware (JTAG) debugger Development Environment
60 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Prototype Performance Tools pmcount • Tool to access to HW performance counters Performance inspector • Suite of GPL based performance analysis tools extended to support SPE threads tprof – timer based analysis tool ptt – per thread time ai – above idle post – report generator a2n – address to name ctrace • Branch tracing performance monitor (under development) Development Environment
61 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation SPE Performance Tools Static analysis (spexlc_timing) • Annotates assembly source with instruction pipeline state Dynamic analysis (CellBE System Simulator) • Generates statistical data on SPE execution Cycles, instructions, and CPI Single/Dual issue rates Stall statistics Register usage Instruction histogramming Development Environment
62 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Miscellaneous Tools – IDL Compiler Development Environment
63 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Samples / Workloads / Demos Numerous code samples provided to demonstrate system design constructs Complex workloads and demos used to evaluate and demonstrate system performance Terrain Rendering Engine Subdivision Surfaces Physics Simulation Geometry Engine Execution Environment
64 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Subsystem Sample – Geometry Engine OpenGL-like geometry engine •Geometry processing is offloaded to compile-time configurable SPE “vertex shader” •User Queue communication model consisting of 4KB blocks for SPU command requests with command headers in SPE Mailbox FIFO PPE Application G E L I B GE Shader User Queues PPE SPE System Memory Viewer Execution Environment
65 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Terrain Rendering Engine Overview Visualization of Terrain Data Increasingly Important •General availability of high resolution satellite images •Publicly available USGS Digital Elevation Models (DEMs) •Mobile GPS devices (land, sea, air) + wireless networks Inferior Current Solutions •Polygonal Models (low quality, non-Real Time) •Requires CPU + GPU (Graphics Processing Unit) Superior CellBE Solution •Highly Compressed Height Map Models (10x less data) •Requires only one CellBE (No GPU) •High Quality Images (Multi-Sampled Raycast) •Fast (Real Time Animations)
66 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Earthviewer – Polygon Render
67 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Real-Time Terrain Rendering Engine (TRE) Cell Optimized Raycaster Advanced SPUAdvanced SPU shadershader functionfunction Real time rendering with only one Cell processoreal time rendering with only one Cell processor No graphics adapter assistNo graphics adapter assist High definition resolutionHigh definition resolution Ray/Terrain intersection computationRay/Terrain intersection computation Texture FilteringTexture Filtering Normal computationNormal computation Bump map computationBump map computation Diffuse + Ambient lighting modelDiffuse + Ambient lighting model PerlinPerlin Noise based cloudsNoise based clouds Atmosphere computation (haze, sun, halo)Atmosphere computation (haze, sun, halo) Dynamic multi-sampling (4Dynamic multi-sampling (4 –– 16 samples per pixel)16 samples per pixel) Image based input (16 bit height + 16 bit texture)Image based input (16 bit height + 16 bit texture) 47 KB of SPU object code47 KB of SPU object code MJPEG like compression via SPUMJPEG like compression via SPU Performance scales linearly with number of availablePerformance scales linearly with number of available SPUsSPUs Written completely in C withWritten completely in C with intrinsicsintrinsics Client supportClient support –– OpenGL workstation, wireless PDAOpenGL workstation, wireless PDA User inputsUser inputs –– graphical, joystick, GPS/accelerometergraphical, joystick, GPS/accelerometer
68 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation 3D Surface from 25x25 Height Data 0 1000 2000 3000 4000 5000 6000 7.5 KB for just the Surface
69 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Raw 25x25 Height Data 0 1000 2000 3000 4000 5000 6000 1.25 KB for Height Map (6x Compression)
70 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Ray Casting
71 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Height Color Data Layout (Main Memory) © 2005 IBM Corporation71 Quad Word (128 bits)Quad Word (128 bits) H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C
72 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation PPE Data Staging via L2 Main MemoryMain Memory PPEPPE L2L2 SPESPE SPESPESPESPE SPESPE SPESPE SPESPESPESPE SPESPE The L2 has 4 Outstanding Loads + 2 Prefetch
73 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation SPE Data Staging 16 Outstanding Loads per SPE Main MemoryMain Memory PPEPPE L2L2 SPESPE SPESPESPESPE SPESPE SPESPE SPESPESPESPE SPESPE
74 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Height Color Data Layout (Main Memory) © 2005 IBM Corporation74 Quad Word (128 bits) H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C Quad Word (128 bits)
75 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Height Color Data Layout (Local Store) © 2005 IBM Corporation75 SIMD Register (128 bits) H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C Shuffle Byte Quad Word (128 bits)
76 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation TRE Image Pipeline Time PPE SPE 0-6 SPE 7 Prep Frame 0Prep Frame 0 Prep Frame 1Prep Frame 1 Render Frame 0Render Frame 0 Render Frame 1Render Frame 1 Encode Frame 0Encode Frame 0 Encode Frame 1Encode Frame 1 Prep Frame 2Prep Frame 2 Prep Frame 3Prep Frame 3 Render Frame 2Render Frame 2
77 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation TRE SPE Render Pipeline MFC Load RC 0 Load RC 0 Load HC/Accm 0Load HC/Accm 0 Load HC/Accm 1Load HC/Accm 1 Store Accm 1 Store Accm 1 Load HC/Accm 0Load HC/Accm 0 Store Accm 0 Store Accm 0 Gen Lists 0Gen Lists 0 Gen Samples 1Gen Samples 1 Gen Lists 0Gen Lists 0Gen Samples 0Gen Samples 0Gen Lists 1Gen Lists 1 Store Accm 0 Store Accm 0 Load RC 0 Load RC 0 Gen Samples 1Gen Samples 1 Load RC 1 Load RC 1 TimeTime SPU
78 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Ray-casting with SIMD Ray 1 Ray 4Ray 3Ray 2 Vector (128 bits) 4 3 2 1
79 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation TRE SPE Ray Kernel Ray IntersectionRay Intersection Multi - Sample AdjustmentMulti - Sample Adjustment Shader Texture Filtering Normal Generation Bump Mapping Lighting Atmosphere Shader Texture Filtering Normal Generation Bump Mapping Lighting Atmosphere Sample AccumulationSample Accumulation
80 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation SPE Local Store Memory Layout Height Color Data (57 KB) Height Color Data (57 KB) Height Color Data (57 KB) Height Color Data (57 KB) Accumulation Data (30 KB) Accumulation Data (30 KB) Accumulation Data (30 KB) Accumulation Data (30 KB) Accum DMA List (15 KB)Accum DMA List (15 KB) Accum DMA List (15 KB)Accum DMA List (15 KB) RCRC RCRC ACAC Stack 4 KBStack 4 KB Code (29 KB) Code (29 KB) HC DMA List (13 KB)HC DMA List (13 KB) HC DMA List (13 KB)HC DMA List (13 KB)
81 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Chip
82 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Sample SPE Simulator Output Performance Cycle count 96187613 Performance Instruction count 113483578 (108381008) Performance CPI 0.85 (0.89) Branch instructions 1255537 Branch taken 826730 Branch not taken 428807 Hint instructions 507711 Hint hit 809520 Single cycle 59886926 ( 62.3%) Dual cycle 24247041 ( 25.2%) Nop cycle 133131 ( 0.1%) Stall due to branch miss 1024965 ( 1.1%) Stall due to prefetch miss 22394 ( 0.0%) Stall due to dependency 9709208 ( 10.1%) Stall due to fp resource conflict 0 ( 0.0%) Stall due to waiting for hint target 1163937 ( 1.2%) Stall due to dp pipeline 0 ( 0.0%) Channel stall cycle 0 ( 0.0%) SPU Initialization cycle 9 ( 0.0%) ----------------------------------------------------------------------- Total cycle 96187611 (100.0%) The number of used registers are 128, the used ratio is 100.00
83 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Austin
84 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Mount Saint Helens
85 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation TRE 720P Performance 2.0 GHz Apple G5 0.6 frames/sec • 40% of cycles spent waiting for Memory 3.2 GHz Cell 30.0 frames/sec • 1% of cycles spent waiting for Memory Cell has 50x advantage
86 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Summary Cell ushers in a new era of leading edge processors optimized for digital media and entertainment Desire for realism is driving a convergence between supercomputing and entertainment New levels of performance and power efficiency beyond what is achieved by PC processors Responsiveness to the human user and the network are key drivers for Cell Cell will enable entirely new classes of applications, even beyond those we contemplate today
87 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation (c) Copyright International Business Machines Corporation 2005. All Rights Reserved. Printed in the United Sates September 2005. The following are trademarks of International Business Machines Corporation in the United States, or other countries, or both. IBM IBM Logo Power Architecture Other company, product and service names may be trademarks or service marks of others. All information contained in this document is subject to change without notice. The products described in this document are NOT intended for use in applications such as implantation, life support, or other hazardous uses where malfunction could result in death, bodily injury, or catastrophic property damage. The information contained in this document does not affect or change IBM product specifications or warranties. Nothing in this document shall operate as an express or implied license or indemnity under the intellectual property rights of IBM or third parties. All information contained in this document was obtained in specific environments, and is presented as an illustration. The results obtained in other operating environments may vary. While the information contained herein is believed to be accurate, such information is preliminary, and should not be relied upon for accuracy or completeness, and no representations or warranties of accuracy or completeness are made. THE INFORMATION CONTAINED IN THIS DOCUMENT IS PROVIDED ON AN "AS IS" BASIS. In no event will IBM be liable for damages arising directly or indirectly from any use of the information contained in this document. IBM Microelectronics Division The IBM home page is http://www.ibm.com 1580 Route 52, Bldg. 504 The IBM Microelectronics Division home page is Hopewell Junction, NY 12533-6351 http://www.chips.ibm.com

More Related Content

PDF
Cell Technology for Graphics and Visualization
Slide_N
 
PPTX
IBM PureSystems
HelpSystems
 
PDF
IBM PureFlex System
IBM India Smarter Computing
 
PDF
Palestra IBM-Mack Zvm linux
Vivaldo Jose Breternitz
 
PDF
Early Successes Debugging with TotalView on the Intel Xeon Phi Coprocessor
Intel IT Center
 
PPTX
Ca lecture 03
Haris456
 
PPTX
System on chip approach
Hamdy Fouad
 
PPSX
System on chip architectures
Dr. A. B. Shinde
 
Cell Technology for Graphics and Visualization
Slide_N
 
IBM PureSystems
HelpSystems
 
IBM PureFlex System
IBM India Smarter Computing
 
Palestra IBM-Mack Zvm linux
Vivaldo Jose Breternitz
 
Early Successes Debugging with TotalView on the Intel Xeon Phi Coprocessor
Intel IT Center
 
Ca lecture 03
Haris456
 
System on chip approach
Hamdy Fouad
 
System on chip architectures
Dr. A. B. Shinde
 

What's hot (19)

PDF
Michael Gschwind, Chip Multiprocessing and the Cell Broadband Engine
Michael Gschwind
 
PDF
Blue Gene Active Storage
Heiko Joerg Schick
 
PDF
Soc - Intro, Design Aspects, HLS, TLM
Subhash Iyer
 
PPTX
IBM BladeCenter Fundamentals Introduction
Dsunte Wilson
 
PDF
System On Chip
Dr. A. B. Shinde
 
PDF
SOC Design Challenges and Practices
Dr. Shivananda Koteshwar
 
PPTX
Ibm aix
mindmajixtrainings
 
PPTX
Intel i7
Justin k.
 
PPTX
Soc lect1
ecemaster
 
PDF
System-on-Chip Design, Embedded System Design Challenges
pboulet
 
PPTX
Intel i7 Technologies
Bibhu Biswal
 
PPTX
System-on-Chip Programmable Retina
Vanya Valindria
 
PDF
System on Chip (SoC) for mobile phones
Jeffrey Funk
 
PDF
POWER9 AC922 Newell System - HPC & AI
Anand Haridass
 
PPTX
What is system on chip (1)
Jagadeshgoud
 
PPT
Power 7 Overview
lambertt
 
PPTX
SOC - system on a chip
Parth Kavi
 
PPT
2123.a better waytoprint.universal print
Sumit Tambe
 
PPTX
Intel core i7 processors
Self employed
 
Michael Gschwind, Chip Multiprocessing and the Cell Broadband Engine
Michael Gschwind
 
Blue Gene Active Storage
Heiko Joerg Schick
 
Soc - Intro, Design Aspects, HLS, TLM
Subhash Iyer
 
IBM BladeCenter Fundamentals Introduction
Dsunte Wilson
 
System On Chip
Dr. A. B. Shinde
 
SOC Design Challenges and Practices
Dr. Shivananda Koteshwar
 
Ibm aix
mindmajixtrainings
 
Intel i7
Justin k.
 
Soc lect1
ecemaster
 
System-on-Chip Design, Embedded System Design Challenges
pboulet
 
Intel i7 Technologies
Bibhu Biswal
 
System-on-Chip Programmable Retina
Vanya Valindria
 
System on Chip (SoC) for mobile phones
Jeffrey Funk
 
POWER9 AC922 Newell System - HPC & AI
Anand Haridass
 
What is system on chip (1)
Jagadeshgoud
 
Power 7 Overview
lambertt
 
SOC - system on a chip
Parth Kavi
 
2123.a better waytoprint.universal print
Sumit Tambe
 
Intel core i7 processors
Self employed
 
Ad

Viewers also liked (20)

PPTX
Chapter 1 engine components and classification
Hafizkamaruddin
 
PPT
Six stroke-engine-presenation
gunjan panchal
 
PPT
The Animated 4 Cycle Internal Combustion Engine
Chris Coggon
 
PPTX
Cryogenic rocket engine
guestfadacb
 
PPT
6 stroke engine
Anand Kannan
 
PPTX
Green engine - an introduction
DIVINE SEBASTIAN
 
KEY
Introduction to Google App Engine
Chakkrit (Kla) Tantithamthavorn
 
PPT
Cloud for Developers: Azure vs. Google App Engine vs. Amazon vs. AppHarbor
Svetlin Nakov
 
PDF
Combustion in diesel engine
Amanpreet Singh
 
PPT
Understanding the Diesel Engine
Biodiesel Automotive
 
PDF
Apache Kylin: OLAP Engine on Hadoop - Tech Deep Dive
Xu Jiang
 
PPT
IC Engine PPt
Rocker ʌvinʌsh
 
ODP
Search Engine Demystified
Sudarsun Santhiappan
 
PDF
RoomCloud Booking Engine
Sylvain Diverchy
 
PPTX
Diesel engine power plant
SURAJ PRASAD
 
PPTX
Hive on spark is blazing fast or is it final
Hortonworks
 
PPT
Fuel Injection and Ignition
guest2c6da6
 
PPT
Search Engine Optimization (SEO)
Dennis Deacon
 
PDF
Periodic Table of SEO Success Factors & Guide to SEO by SearchEngineLand
Search Engine Land
 
PPT
Ic engine
NFC Institute of engineering and fertilizer research Faisalabad
 
Chapter 1 engine components and classification
Hafizkamaruddin
 
Six stroke-engine-presenation
gunjan panchal
 
The Animated 4 Cycle Internal Combustion Engine
Chris Coggon
 
Cryogenic rocket engine
guestfadacb
 
6 stroke engine
Anand Kannan
 
Green engine - an introduction
DIVINE SEBASTIAN
 
Introduction to Google App Engine
Chakkrit (Kla) Tantithamthavorn
 
Cloud for Developers: Azure vs. Google App Engine vs. Amazon vs. AppHarbor
Svetlin Nakov
 
Combustion in diesel engine
Amanpreet Singh
 
Understanding the Diesel Engine
Biodiesel Automotive
 
Apache Kylin: OLAP Engine on Hadoop - Tech Deep Dive
Xu Jiang
 
IC Engine PPt
Rocker ʌvinʌsh
 
Search Engine Demystified
Sudarsun Santhiappan
 
RoomCloud Booking Engine
Sylvain Diverchy
 
Diesel engine power plant
SURAJ PRASAD
 
Hive on spark is blazing fast or is it final
Hortonworks
 
Fuel Injection and Ignition
guest2c6da6
 
Search Engine Optimization (SEO)
Dennis Deacon
 
Periodic Table of SEO Success Factors & Guide to SEO by SearchEngineLand
Search Engine Land
 
Ic engine
NFC Institute of engineering and fertilizer research Faisalabad
 
Ad

Similar to Hardware and Software Architectures for the CELL BROADBAND ENGINE processor (20)

PDF
Cell Today and Tomorrow - IBM Systems and Technology Group
Slide_N
 
PDF
Future Commodity Chip Called CELL for HPC
Slide_N
 
PDF
M. Gschwind, A novel SIMD architecture for the Cell heterogeneous chip multip...
Michael Gschwind
 
PDF
Michael Gschwind et al, "An Open Source Environment for Cell Broadband Engine...
Michael Gschwind
 
PDF
Synergistic processing in cell's multicore architecture
Michael Gschwind
 
PDF
Cell Broadband EngineTM: and Cell/B.E. based blade technology
Slide_N
 
PPT
Xtw01t2v012011 sys tech
pgnguyen44
 
PPT
The Cell Processor
Heiko Joerg Schick
 
PDF
IBM: Introduction to the Cell Broadband Engine Architecture
Slide_N
 
PDF
Deeplearningusingcloudpakfordata
Ganesan Narayanasamy
 
PDF
Toshiba's Approach to Consumer Product Applications by Cell and Desire/Challe...
Slide_N
 
PDF
Michael Gschwind, Cell Broadband Engine: Exploiting multiple levels of parall...
Michael Gschwind
 
PDF
OpenPOWER Seminar at IIT Madras
Ganesan Narayanasamy
 
PDF
OpenPOWER Seminar at IIIT Bangalore
Ganesan Narayanasamy
 
PDF
Chip Multiprocessing and the Cell Broadband Engine.pdf
Slide_N
 
PDF
IBM POWER8: The first OpenPOWER processor
Slide_N
 
PPT
Ibm cell
Vyanktesh Dorlikar
 
PPT
STG101 Power Product_PRESENTATION VERSION 1.ppt
salesjpayroll
 
PDF
Industry Trends in Microprocessor Design
Slide_N
 
PPT
Unisanta - Visão Geral de hardware Servidor IBM System z
Anderson Bassani
 
Cell Today and Tomorrow - IBM Systems and Technology Group
Slide_N
 
Future Commodity Chip Called CELL for HPC
Slide_N
 
M. Gschwind, A novel SIMD architecture for the Cell heterogeneous chip multip...
Michael Gschwind
 
Michael Gschwind et al, "An Open Source Environment for Cell Broadband Engine...
Michael Gschwind
 
Synergistic processing in cell's multicore architecture
Michael Gschwind
 
Cell Broadband EngineTM: and Cell/B.E. based blade technology
Slide_N
 
Xtw01t2v012011 sys tech
pgnguyen44
 
The Cell Processor
Heiko Joerg Schick
 
IBM: Introduction to the Cell Broadband Engine Architecture
Slide_N
 
Deeplearningusingcloudpakfordata
Ganesan Narayanasamy
 
Toshiba's Approach to Consumer Product Applications by Cell and Desire/Challe...
Slide_N
 
Michael Gschwind, Cell Broadband Engine: Exploiting multiple levels of parall...
Michael Gschwind
 
OpenPOWER Seminar at IIT Madras
Ganesan Narayanasamy
 
OpenPOWER Seminar at IIIT Bangalore
Ganesan Narayanasamy
 
Chip Multiprocessing and the Cell Broadband Engine.pdf
Slide_N
 
IBM POWER8: The first OpenPOWER processor
Slide_N
 
Ibm cell
Vyanktesh Dorlikar
 
STG101 Power Product_PRESENTATION VERSION 1.ppt
salesjpayroll
 
Industry Trends in Microprocessor Design
Slide_N
 
Unisanta - Visão Geral de hardware Servidor IBM System z
Anderson Bassani
 

More from Slide_N (20)

PDF
IBM: Introduction to the Cell Multiprocessor
Slide_N
 
PDF
AMD: The Next Generation of Microprocessors
Slide_N
 
PDF
Cryptologic Applications of the PlayStation 3: Cell SPEED
Slide_N
 
PDF
Toward an Open and Unified Model for Heterogeneous and Accelerated Multicore ...
Slide_N
 
PDF
Roadrunner: Heterogeneous Petascale Computing for Predictive Simulation
Slide_N
 
PDF
Driving a Hybrid in the Fast-lane: The Petascale Roadrunner System at Los Alamos
Slide_N
 
PDF
Petascale Visualization: Approaches and Initial Results
Slide_N
 
PDF
The Cell at Los Alamos: From Ray Tracing to Roadrunner
Slide_N
 
PDF
Roadrunner and hybrid computing - Conference on High-Speed Computing
Slide_N
 
PDF
Roadrunner Tutorial: An Introduction to Roadrunner and the Cell Processor
Slide_N
 
PDF
Deferred Pixel Shading on the PlayStation 3
Slide_N
 
PDF
POWER9: IBM’s Next Generation POWER Processor
Slide_N
 
PDF
IBM POWER8 Systems Technology Group Development
Slide_N
 
PDF
Efficient Usage of Compute Shaders on Xbox One and PS4
Slide_N
 
PDF
Common Software Models and Platform for Cell and SpursEngine
Slide_N
 
PDF
Towards Cell Broadband Engine - Together with Playstation
Slide_N
 
PDF
SpursEngine A High-performance Stream Processor Derived from Cell/B.E. for Me...
Slide_N
 
PDF
Parallel Vector Tile-Optimized Library (PVTOL) Architecture-v3.pdf
Slide_N
 
PDF
Experiences with PlayStation VR - Sony Interactive Entertainment
Slide_N
 
PDF
SPU-based Deferred Shading for Battlefield 3 on Playstation 3
Slide_N
 
IBM: Introduction to the Cell Multiprocessor
Slide_N
 
AMD: The Next Generation of Microprocessors
Slide_N
 
Cryptologic Applications of the PlayStation 3: Cell SPEED
Slide_N
 
Toward an Open and Unified Model for Heterogeneous and Accelerated Multicore ...
Slide_N
 
Roadrunner: Heterogeneous Petascale Computing for Predictive Simulation
Slide_N
 
Driving a Hybrid in the Fast-lane: The Petascale Roadrunner System at Los Alamos
Slide_N
 
Petascale Visualization: Approaches and Initial Results
Slide_N
 
The Cell at Los Alamos: From Ray Tracing to Roadrunner
Slide_N
 
Roadrunner and hybrid computing - Conference on High-Speed Computing
Slide_N
 
Roadrunner Tutorial: An Introduction to Roadrunner and the Cell Processor
Slide_N
 
Deferred Pixel Shading on the PlayStation 3
Slide_N
 
POWER9: IBM’s Next Generation POWER Processor
Slide_N
 
IBM POWER8 Systems Technology Group Development
Slide_N
 
Efficient Usage of Compute Shaders on Xbox One and PS4
Slide_N
 
Common Software Models and Platform for Cell and SpursEngine
Slide_N
 
Towards Cell Broadband Engine - Together with Playstation
Slide_N
 
SpursEngine A High-performance Stream Processor Derived from Cell/B.E. for Me...
Slide_N
 
Parallel Vector Tile-Optimized Library (PVTOL) Architecture-v3.pdf
Slide_N
 
Experiences with PlayStation VR - Sony Interactive Entertainment
Slide_N
 
SPU-based Deferred Shading for Battlefield 3 on Playstation 3
Slide_N
 

Recently uploaded (20)

PDF
Microsoft Solutions Partner Drive Digital Transformation with D365.pdf
Microsoft Dynamics
 
PPTX
observCloud-Native Containerability and monitoring.pptx
anupsingh76937
 
PDF
Zenith AI: Advanced Artificial Intelligence
Biz Preneurs
 
PDF
A contest of sentiment analysis: k-nearest neighbor versus neural network
IAESIJAI
 
PDF
A novel scalable deep ensemble learning framework for big data classification...
IAESIJAI
 
PDF
Transform Your ITIL® 4 & ITSM Strategy with AI in 2025.pdf
PDCA Consulting
 
PDF
1 - Historical Antecedents, Social Consideration.pdf
tuazon2030627
 
PPTX
O2C Customer Invoices to Receipt V15A.pptx
ssuserbfa915
 
PDF
NewMind AI Weekly Chronicles – August ’25 Week III
NewMind AI
 
PPT
Module 1.ppt Iot fundamentals and Architecture
nanditha7766
 
PPT
Geologic Time for studying geology for geologist
ssuser738f5a
 
PPTX
MicrosoftCybserSecurityReferenceArchitecture-April-2025.pptx
SilvioHayashi
 
PDF
Unlock new opportunities with location data.pdf
Precisely
 
PDF
STKI Israel Market Study 2025 version august
Dr. Jimmy Schwarzkopf
 
PDF
Five Habits of High-Impact Board Members
OnBoard
 
PDF
Hybrid horned lizard optimization algorithm-aquila optimizer for DC motor
IAESIJAI
 
PPT
What is a Computer? Input Devices /output devices
GulJan8
 
PDF
DASA ADMISSION 2024_FirstRound_FirstRank_LastRank.pdf
rameswartudu05
 
PPTX
Chapter 5: Probability Theory and Statistics
RandyFin
 
PDF
Developing a website for English-speaking practice to English as a foreign la...
IAESIJAI
 
Microsoft Solutions Partner Drive Digital Transformation with D365.pdf
Microsoft Dynamics
 
observCloud-Native Containerability and monitoring.pptx
anupsingh76937
 
Zenith AI: Advanced Artificial Intelligence
Biz Preneurs
 
A contest of sentiment analysis: k-nearest neighbor versus neural network
IAESIJAI
 
A novel scalable deep ensemble learning framework for big data classification...
IAESIJAI
 
Transform Your ITIL® 4 & ITSM Strategy with AI in 2025.pdf
PDCA Consulting
 
1 - Historical Antecedents, Social Consideration.pdf
tuazon2030627
 
O2C Customer Invoices to Receipt V15A.pptx
ssuserbfa915
 
NewMind AI Weekly Chronicles – August ’25 Week III
NewMind AI
 
Module 1.ppt Iot fundamentals and Architecture
nanditha7766
 
Geologic Time for studying geology for geologist
ssuser738f5a
 
MicrosoftCybserSecurityReferenceArchitecture-April-2025.pptx
SilvioHayashi
 
Unlock new opportunities with location data.pdf
Precisely
 
STKI Israel Market Study 2025 version august
Dr. Jimmy Schwarzkopf
 
Five Habits of High-Impact Board Members
OnBoard
 
Hybrid horned lizard optimization algorithm-aquila optimizer for DC motor
IAESIJAI
 
What is a Computer? Input Devices /output devices
GulJan8
 
DASA ADMISSION 2024_FirstRound_FirstRank_LastRank.pdf
rameswartudu05
 
Chapter 5: Probability Theory and Statistics
RandyFin
 
Developing a website for English-speaking practice to English as a foreign la...
IAESIJAI
 

Hardware and Software Architectures for the CELL BROADBAND ENGINE processor

  • 1. IBM Systems and Technology Group © 2005 IBM Corporation Tutorial Hardware and Software Architectures for the CELL BROADBAND ENGINE processor Michael Day, Peter Hofstee IBM Systems & Technology Group, Austin, Texas CODES+ISSS Conference, September 2005
  • 2. 2 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Agenda Trends in Processors and Systems Introduction to Cell Challenges to processor performance Cell Broadband Engine Architecture (CBEA) Cell Broadband Engine (CBE) Processor Overview Cell Programming Models Prototype Software Environment TRE Demo © 2005 IBM Corporation
  • 3. 3 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Trends in Microprocessors and Systems
  • 4. 4 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Processor Performance over Time (Game processors take the lead on media performance) 10 100 1000 10000 100000 1000000 1993 1995 1996 1998 2000 2002 2004 Game Processors PC Processors Flops (SP)
  • 5. 5 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation System Trends toward Integration Implied loss of system configuration flexibility Must compensate with generality of acceleration function to maintain market size. Processor Southbridge NorthbridgeMemory Accel Next-Gen Processor Memory IO IO
  • 6. 6 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Motivation: Cell Goals Outstanding performance, especially on game/multimedia applications. • Challenges: Memory Wall, Power Wall, Frequency Wall Real time responsiveness to the user and the network. • Challenges: Real-time in an SMP environment, Security Applicable to a wide range of platforms. • Challenge: Maintain programmability while increasing performance Support an introduction in 2005/6. • Challenge: Structure innovation such that 5yr. schedule can be met
  • 7. 7 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Performance Limiters in Conventional Microprocessors Memory Wall • Latency induced bandwidth limitations Power Wall • Must improve efficiency and performance equally Frequency Wall • Diminishing returns from deeper pipelines (can be negative if power is taken into account)
  • 8. 8 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell © 2005 IBM Corporation Systems and Technology Group
  • 9. 9 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell History IBM, SCEI/Sony, Toshiba Alliance formed in 2000 Design Center opened in March 2001 Based in Austin, Texas February 7, 2005: First external technical disclosures • Cell Broadband Engine Architecture documentation can be found at: http://www.ibm.com/developerworks/power/cell • Additional publications on Cell can be downloaded from: http://www.ibm.com/chips/techlib/techlib.nsf/products/Cell • A paper on Cell in the upcoming issue of the IBM Journal of Research and Development can be found at: http://www.research.ibm.com/journal/rd/494/kahle.html Systems and Technology Group © 2005 IBM Corporation
  • 10. 10 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Highlights Supercomputer on a chip Multi-core microprocessor (9 cores) 3.2 GHz clock frequency 10x performance for many applications Digital home to distributed computing Systems and Technology Group © 2005 IBM Corporation
  • 11. 11 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Introducing Cell Sets a new performance standard • Exploits parallelism while achieving high frequency • Supercomputer attributes with extreme floating point capabilities • Sustains high memory bandwidth with smart DMA controllers Designed for natural human interaction • Photo-realistic effects • Predictable real-time response • Virtualized resources for concurrent activities Designed for flexibility • Wide variety of application domains • Highly abstracted to highly exploitable programming models • Reconfigurable I/O interfaces • Autonomic power management
  • 12. 12 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Microprocessor Architecture Trends CPU GPU NIC Security Media … CPU Streaming Graphics Processor Network Processor Security Processor Media Processor … 64b Power Processor Synergistic Processor Synergistic Processor ... Mem. Contr. Config. IO Hardwired Function Programmable ASIC Cell
  • 13. 13 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Key Attributes of Cell Cell is Multi-Core • Contains 64-bit Power Architecture TM • Contains 8 Synergistic Processor Elements (SPE) Cell is a Flexible Architecture • Multi-OS support (including Linux) with Virtualization technology • Path for OS, legacy apps, and software development Cell is a Broadband Architecture • SPE is RISC architecture with SIMD organization and Local Store • 128+ concurrent transactions to memory per processor Cell is a Real-Time Architecture • Resource allocation (for Bandwidth Measurement) • Locking Caches (via Replacement Management Tables) Cell is a Security Enabled Architecture • SPE dynamically reconfigurable as secure processors
  • 14. 14 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Architecture is … COHERENT BUS Power ISA MMU/BIU Power ISA MMU/BIU … IO transl. Memory Incl. coherence/memory compatible with 32/64b Power Arch. Applications and OS’s 64b Power Architecture™
  • 15. 15 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Architecture is … 64b Power Architecture™ COHERENT BUS (+RAG) Power ISA +RMT MMU/BIU +RMT Power ISA +RMT MMU/BIU +RMT IO transl. Memory Plus Memory Flow Control (MFC) MMU/DMA +RMT Local Store Memory MMU/DMA +RMT Local Store Memory LS Alias LS Alias … … …
  • 16. 16 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Architecture is … 64b Power Architecture™+ MFC COHERENT BUS (+RAG) Power ISA +RMT MMU/BIU +RMT Power ISA +RMT MMU/BIU +RMT IO transl. Memory Plus Synergistic Processors MMU/DMA +RMT Local Store Memory MMU/DMA +RMT Local Store Memory LS Alias LS Alias … … … Syn. Proc. ISA Syn. Proc. ISA
  • 17. 17 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell - Attacking the Performance Walls Multi-Core Non-Homogeneous Architecture • Control Plane vs. Data Plane processors • Attacks Power Wall 3-level Model of Memory • Main Memory, Local Store, Registers • Attacks Memory Wall Large Shared Register File & SW Controlled Branching • Allows deeper pipelines • Attacks Frequency Wall Systems and Technology Group © 2005 IBM Corporation
  • 18. 18 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Frequency Increase vs Power Consumption 0 0.5 1 1.5 2 2.5 3 3.5 0.9 1 1.1 1.2 1.3 Voltage Realative Power Frequency
  • 19. 19 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Power Efficient Architecture and the CellBE Non-Homogeneous Coherent Multi-Processor • Data-plane/Control-plane specialization • More efficient than homogeneous SMP 3-level model of Memory • Bandwidth without (inefficient) speculation • High-bandwidth .. Low power
  • 20. 20 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Power Efficient Architecture and the SPE Power Efficient ISA allows Simple Control • Single mode architecture • No cache • Branch hint • Large unified register file • Channel Interface Efficient Microarchitecture • Single port local store • Extensive clock gating Efficient implementation • See Cool Chips paper by O. Takahashi et al. and T. Asano et al.
  • 21. 21 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Broadband Engine Components
  • 22. 22 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Processor Components Power Core (PPE) L2 Cache NCU Power Processor Element (PPE): •General Purpose, 64-bit RISC Processor (PowerPC AS 2.0.2) •2-Way Hardware Multithreaded •L1 : 32KB I ; 32KB D •L2 : 512KB •Coherent load/store •VMX •3+ GHz •Real-time Control •Locking L2 Cache & TLB •Bandwidth Reservation In the Beginning – the solitary Power Processor Custom Designed – for high frequency, space and power efficiency
  • 23. 23 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation 96 Byte/Cycle 300+GB/sec @ 3.2 GHz Element Interconnect Bus Cell Processor Components Power Core (PPE) L2 Cache NCU N N N MIC EIB data ring for internal communication • Four 16 byte data rings, supporting multiple transfers • 96B/cycle peak bandwidth • Over 100 outstanding requests • 300+ GByte/sec @ 3.2 GHz
  • 24. 24 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation 96 Byte/Cycle 300+ GB/sec @ 3.2 GHz Element Interconnect Bus Cell Processor Components Power Core (PPE) L2 Cache NCU N N N MIC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC Local Store SPUMFC N AUC Local Store SPUMFC N AUC SPE provides computational performance • Dual issue, up to 8-way 32-bit SIMD • Dedicated resources: 128 128-bit RF, 256KB Local Store • Each DMA MMU can be dynamically configured to protect resources • Dedicated DMA engine: Up to 16 outstanding requests Local Store SPU MFC N AUC Local Store SPU MFC N AUC
  • 25. 25 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation 96 Byte/Cycle 300+ GB/sec @ 3.2GHz Element Interconnect Bus Cell Processor Components Power Core (PPE) L2 Cache NCU N N N MIC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC Local Store SPUMFC N AUC Local Store SPUMFC N AUC Memory Flow Controller (MFC) and Atomic Update Cache (AUC) • 4 line cache for shared memory atomic update primitives •High performance DMA unit •Local Store aliased into PPE system memory •16 element SPE side DMA Command Queue • 8 element PPE side DMA Command Queue •DMA List supports 2K entry scatter/gather •MFC-MMU controls SPE DMA accesses •Compatible with PowerPC Virtual Memory architecture •Full memory protection for MFC DMA •S/W controllable from PPE MMIO •DMA 1,2,4,8,16,128 -> 16Kbyte transfers for I/O access
  • 26. 26 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Processor Components IOIF0 20 GB/sec BIF or IOIF0 IOIF1 Southbridge I/O 5 GB/sec LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC Local Store SPUMFC N AUC Local Store SPUMFC N AUC Local Store SPU MFC N AUC Local Store SPU MFC N AUC 96 Byte/Cycle Element Interconnect Bus Power Core (PPE) L2 Cache NCU Memory Interface Controller (MIC): • Dual XDRTM controller (25.6GB/s @ 3.2Gbps) • ECC support • Suspend to DRAM support Broadband Interface Controller (BIC): Provides a wide connection to external devices Two configurable interfaces (50+GB/s @ 5Gbps) – Configurable number of bytes – Coherent (BIF) and / or I/O (IOIFx) protocols Supports two virtual channels per interface Supports multiple system configurations MIC 25 GB/sec XDR DRAM
  • 27. 27 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Processor Components I/O Bus Master Translation (IOT) Translates Bus Addresses to System Real Addresses Two Level Translation – I/O Segments (256 MB) – I/O Pages (4K, 64K, 1M, 16M byte) I/O Device Identifier per page for LPAR IOST and IOPT Cache – hardware / software managed IOIF0 20 GB/sec BIF or IOIF0 MIC 25 GB/sec XDR DRAM IOIF1 Southbridge I/O 5 GB/sec LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC Local Store SPUMFC N AUC Local Store SPUMFC N AUC Local Store SPU MFC N AUC Local Store SPU MFC N AUC 96 Byte/Cycle Element Interconnect Bus Power Core (PPE) L2 Cache NCU IIC IOT Internal Interrupt Controller (IIC) Handles SPE Interrupts Handles External Interrupts – From Coherent Interconnect – From IOIF0 or IOIF1 Interrupt Priority Level Control Interrupt Generation Ports for IPI Duplicated for each PPE hardware thread
  • 28. 28 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation 96 Byte/Cycle 256GB/sec @ 4GHz Element Interconnect Bus Cell Processor Components Power Core (PPE) L2 Cache NCU N N LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC Local Store SPUMFC N AUC Local Store SPUMFC N AUC LocalStore SPUMFC N AUC LocalStore SPUMFC N AUC MIC IOIF1 IOIF0 Southbridge I/O 20 GB / sec BIF or IOIF1 25 GB / sec XDR DRAM 5 GB/sec Token Manager – Bandwidth Reservation for shared resources • Optionally Enabled for RT Tasks or LPAR • Multiple Resource Allocation Groups • Generates access tokens at configurable rate for each allocation group 1 per each memory bank (16 total) 2 for each IOIF (4 total) • Requestors assigned RAG ID by OS/hypervisor Each SPE PPE L2 / NCU IOIF 0 Bus Master IOIF 1 Bus Master • Priority order for using another RAGs unused tokens • Resource overcommit warning interrupt IIC IOTTKM
  • 29. 29 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell ProcessorCell Processor Power Processor Element (PPE): •General Purpose, 64-bit RISC Processor (PowerPC AS 2.0) •2-Way Hardware Multithreaded •L1 : 32KB I ; 32KB D •L2 : 512KB •Coherent load/store •VMX •3.2 GHz Internal Interconnect: •Coherent ring structure •300+ GB/s total internal interconnect bandwidth •DMA control to/from SPEs supports >100 outstanding memory requests Synergistic Processor Elements (SPE): •8 per chip •128-bit wide SIMD Units •Integer and Floating Point capable •256KB Local Store •Up to 25.6 GF/s per SPE --- 200GF/s total * External Interconnects: •25.6 GB/sec BW memory interface •2 Configurable I/O Interfaces •Coherent interface (SMP) •Normal I/O interface (I/O & Graphics) •Total BW configurable between interfaces •Up to 35 GB/s out •Up to 25 GB/s in Memory Management & Mapping •SPE Local Store aliased into PPE system memory •MFC/MMU controls SPE DMA accesses •Compatible with PowerPC Virtual Memory architecture •S/W controllable from PPE MMIO •Hardware or Software TLB management •SPE DMA access protected by MFC/MMU ““Supercomputer-on-a-Chip” * At clock speed of 3.2GHz Element Interconnect Bus MFC Local Store SPU N AUC MFC Local Store SPU N AUC Power Core (PPE) L2 Cache NCU Local Store SPU MFC N AUC Local Store SPU MFC N AUC N N N N MFC LocalStore SPU N AUC MFC LocalStore SPU N AUC MFC LocalStore SPU N AUC MFC LocalStore SPU N AUC 20 GB/sec Coherent Interconnect 25.6 GB/sec Memory Inteface 5 GB/sec I/O Bus
  • 30. 30 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation SPE Highlights RISC like organization • 32 bit fixed instructions • Clean design – unified Register file User-mode architecture • No translation/protection within SPU • DMA is full Power Arch protect/x-late VMX-like SIMD dataflow • Broad set of operations (8 / 16 / 32 Byte) • Graphics SP-Float • IEEE DP-Float Unified register file • 128 entry x 128 bit 256KB Local Store • Combined I & D • 16B/cycle L/S bandwidth • 128B/cycle DMA bandwidth LS LS LS LS GPR FXU ODD FXU EVN SFP DP CONTROL CHANNEL DMA SMM ATO SBI RTB FWD 14.5mm2 (90nm SOI)
  • 31. 31 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation What is a Synergistic Processor? (and why is it efficient?) Local Store “is” large 2nd level register file / private instruction store instead of cache • Asynchronous transfer (DMA) to shared memory • Frontal attack on the Memory Wall Media Unit turned into a Processor • Unified (large) Register File • 128 entry x 128 bit Media & Compute optimized • One context • SIMD architecture LS LS LS LS GPR FXU ODD FXU EVN SFP DP CONTROL CHANNEL DMA SMM ATO SBI RTB FWD SPU SMF
  • 32. 32 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation SPE Detail LS LS LS LS GPR FXU ODD FXU EVN SFP DP CONTROL CHANNEL DMA SMM ATO SBI RTB FWD Synergistic Processor Element (SPE) User-mode architecture • No translation/protection within SPE • DMA is full PowerPC protect/xlate Direct programmer control • DMA/DMA-list • Branch hint VMX-like SIMD dataflow • Graphics SP-Float • No saturate arith, some byte • IEEE DP-Float (BlueGene- like) Unified register file • 128 entry x 128 bit 256KB Local Store • Combined I & D • 16B/cycle L/S bandwidth • 128B/cycle DMA bandwidth SPE Latencies • Simple fixed point - 2 cycles* • Complex fixed point - 4 cycles* • Load - 6 cycles* Local store size = 256 KB • Single-precision (ER) float - 6 cycles* • Integer multiply - 7 cycles* • Branch miss - 20 cycles No penalty if correctly hinted • DP (IEEE) float - 13 cycles* Partially pipelined • Enqueue DMA Command - 20 cycles* SPU Units: • Simple (FXU even) Add/Compare Rotate Logical, Count Leading Zero • Permute (FXU odd) Permute Table-lookup • FPU (Single / Double Precision) • Control (SCN) Dual Issue, Load/Store, ECC Handling • Channel (SSC) – Interface to MFC • Register File (GPR/FWD)
  • 33. 33 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Coherent Offload Model DMA into and out of Local Store similar to Power core loads & stores Governed by Power Architecture page and segment tables for translation and protection Shared memory model • Power architecture compatible addressing • MMIO capabilities for SPEs • Local Store is mapped (alias) allowing LS to LS DMA transfers • DMA equivalents of locking loads & stores • OS management/virtualization of SPEs Pre-emptive context switch is supported
  • 34. 34 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation MFC Detail Memory Flow Control System •DMA Unit •LS <-> LS, LS<-> Sys Memory, LS<-> I/O Transfers •8 PPE-side Command Queue entries •16 SPU-side Command Queue entries •MMU similar to PowerPC MMU •8 SLBs, 256 TLBs •multiple page sizes •Software/HW page table walk •PT/SLB misses interrupt PPE •Atomic Cache Facility •4 cache lines for atomic updates •2 cache lines for cast out/MMU reload •Up to 16 outstanding DMA requests in BIU •Resource / Bandwidth Management Tables •Token Based Bus Access Management •TLB Locking Isolation Mode Support (Security Feature) Hardware enforced “isolation” • SPE and Local Store not visible (bus or jtag) • Small LS “untrusted area” for communication area Secure Boot • Chip Specific Key • Decrypt/Authenticate Boot code “Secure Vault” – Runtime Isolation Support • Isolate Load Feature • Isolate Exit Feature Legend: Data Bus Snoop Bus Control Bus Xlate Ld/St MMIO Local Store SPU DMA Engine DMA Queue Atomic Facility MMU RMT Bus I/F Control SPE MMIO
  • 35. 35 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Implementation Aspects
  • 36. 36 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Pre-Decode L1 Instruction Cache Microcode SMT Dispatch (Queue) Decode Dependency Issue Branch Scan Fetch Control L2 Interface VMX/FPU Issue (Queue) VMX Load/Store/ Permute VMX Arith./Logic Unit FPU Load/Store FPU Arith/Logic Unit Load/Store Unit Branch Execution Unit Fixed-Point Unit FPU CompletionVMX Completion Completion/Flush 8 4 2 1 2 11 1 1 111 4 Thread A Thread B Threads alternate fetch and dispatch cycles Thread A Thread B Thread A L1 Data Cache 2 PPE BLOCK DIAGRAM
  • 37. 37 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation ID2 IS1 RF2 EX1 EX2 IC1 IC2 IC3 ID1 RF1 RF2 EX1 EX3 EX4EX2 MC1 IS2 RF1 RF2 EX1 EX2 EX3 EX4 Fixed Point Unit Instruction PPE PIPELINE FRONT END PPE PIPELINE BACK END Microcode MC2 MC3 MC4 EX5 EX5 ... MC9 MC10 MC11 IB2 IS3 EX7 WB EX6 IB1 ID3 EX3RF1DLY DLY DLY DLY DLY DLY EX4 IBZ IC0 Branch Instruction EX8 WB IC4 BP1 BP2 BP3 BP4 Load/Store Instruction Branch Prediction Instruction Decode and IssueInstruction Cache and Buffer IC Instruction Cache IB Instruction Buffer BP Branch Prediction MC Microcode ID Instruction Decode IS Instruction Issue DLY Delay Stage RF Register File Access EX Execution WB Write Back
  • 38. 38 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Permute Unit Load-Store Unit Floating-Point Unit Fixed-Point Unit Branch Unit Channel Unit Result Forwarding and Staging Register File Local Store (256kB) Single Port SRAM 128B Read 128B Write DMA Unit Instruction Issue Unit / Instruction Line Buffer 8 Byte/Cycle 16 Byte/Cycle 128 Byte/Cycle64 Byte/Cycle On-Chip Coherent Bus SPE BLOCK DIAGRAM
  • 39. 39 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation EX1 EX3 EX4EX2 SPU PIPELINE FRONT END SPE PIPELINE BACK END EX5 EX6 RF1 RF2 Branch Instruction WB Load/Store Instruction IF Instruction Fetch IB Instruction Buffer ID Instruction Decode IS Instruction Issue RF Register File Access EX Execution WB Write Back IF1 IF2 ID2 IS1IF3 IF4 IF5 ID1 IS2IB2IB1 ID3 EX2 Fixed Point Instruction WBEX1 Floating Point Instruction WBEX1 EX2 Permute Instruction WBEX1 EX3 EX4 EX5 EX6EX2 EX3 EX4
  • 40. 40 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation CellBE Processor ~250M transistors ~235mm2 Top frequency >3GHz 9 cores, 10 threads > 200+ GFlops (SP) @3.2 GHz > 20+ GFlops (DP) @3.2 GHz Up to 25.6GB/s memory B/W Up to 50+ GB/s I/O B/W ~400M$(US) design investment
  • 41. 41 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Processor Can Support Many Systems Game console systems Blades HDTV Home media servers Supercomputers Cell BE Processor XDRtm XDRtm IOIF0 IOIF1 Cell BE Processor XDRtm XDRtm IOIF BIF Cell BE Processor XDRtm XDRtm IOIF Cell BE Processor XDRtm XDRtm IOIF BIF Cell BE Processor XDRtm XDRtm IOIF CellBE Processor XDRtm XDRtm IOIF BIF CellBE Processor XDRtm XDRtm IOIFSW
  • 42. 42 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell BE Processor Initial Application Areas Cell excels at processing of rich media content in the context of broad connectivity • Digital content creation (games and movies) • Game playing and game serving • Distribution of dynamic, media rich content • Imaging and image processing • Image analysis (e.g. video surveillance) • Next-generation physics-based visualization • Video conferencing • Streaming applications (codecs etc.) • Physical simulation & science Cell is an excellent match for any applications that require: • Parallel processing • Real time processing • Graphics content creation or rendering • Pattern matching • High-performance SIMD capabilities
  • 43. 43 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Programmability Programmer Productivity Raw Hardware Performance BE Design 25 GB/sec Memory Inteface 20 GB/sec Coherent Interconnect 5 GB/sec I/O Bus Cell Software Overview 256 GB/sec Coherent Ring MFC Local Store SPU N AUC MFC Local Store SPU N AUC Power Processor (PPE) L2 Cache NCU Local Store SPU MFC N AUC Local Store SPU MFC N AUC N N N N Cell Chip MFC LocalStore SPU N AUC MFC LocalStore SPU N AUC MFC LocalStore SPU N AUC MFC LocalStore SPU N AUC Cell Programming Features Flexible Programming Models Function Offload Model Existing Application Accelerator Parallel Computational Acceleration Heterogeneous Multi-Threading Runtime
  • 44. 44 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Programming Characteristics Exploit all of Cell’s 18 asynchronous engines • Through function offload, or parallel computational tasks Decompose work into data parallel blocks • Independent Data parallel tasks are most efficient Overlap SPE compute with DMA loads and stores • Multibuffering , pipelining Reduce PPE/SPE workload ratio • PPE as Control / OS processor • SPE as heavy lifting computational engines Super-Linear speedups over conventional processors may be achieved • Through exploitation of asynchronous DMA engines
  • 45. 45 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Features Exploited by Software Keeping Intermediate/Control Data on-Chip MMU – ERATS, SLBs, TLBs DMA from L2 cache-> LS LS to LS DMA Cache <-> Cache transfers (atomic update) SPE Signalling Registers SPE <-> PPE Mailboxes Resource Reservation and Allocation PPE can be shared across logical partitions SPEs can be assigned to logical partitions SPEs independently or Group Allocated Cache Replacement Management TLB Replacement Management Bandwidth Reservation Cell Programming Features Power Processor (PPE) L2 Cache MFC Local Store SPU N AUC DMA with Intervention Atomic Update Cache Local Store to Local Store DMA Hardware Managed Cache Coherency Cell Chip System Memory I/O MFC Local Store SPU N AUC MFC Local Store SPU N AUC MFC Local Store SPU N AUC MFC Local Store SPU N AUC MFC Local Store SPU N AUC BIF/IOIF
  • 46. 46 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Subsystem Programming Model Dedicated Function (problem/privileged subsystem) Programmer writes/uses SPU "libraries" Graphics Pipeline Audio Processing MPEG Encoding/Decoding Encryption / Decryption Main Application in PPE, invokes SPU bound services RPC Like Function Call I/O Device Like Interface (FIFO/ Command Queue) 1 or more SPUs cooperating in subsystem Problem State (Application Allocated) –Transcoding –Realtime data transformation & streaming –Graphics Processing –MPEG Encoding / Decoding –Physical Simulation Privileged State (OS Allocated) –Encryption / Decryption Services –Network Packet Filtering / Routing Code-to-data or data-to-code pipelining possible Very efficient in real-time data streaming applications Function Offload Power Core (PPE) Sys. Memory/QoS SPU Local Store MFC N SPU Local Store MFC N Multi-stage Pipeline SPU Local Store MFC N SPU Local Store MFC N SPU Local Store MFC N Parallel-stages Power Core (PPE) Sys. Memory/QoS
  • 47. 47 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Application Specific Acceleration ModelApplication Specific Acceleration Model –– SPC Accelerated SubsystemsSPC Accelerated Subsystems System Memory Parameter Area PPE PowerPC Application Cell Aware OS (Linux) Compression/ Decompression Realtime MPEG Encoding TCP Offload Data Encryption Data Decryption Conventional PowerPC App. with Library Dependencies OpenGL Encrypt Decrypt Encoding OS loads 3rd party function libraries containing PPC wrapped SPU code PPE part of library code loads SPU allocated by OS Loads SPU part of library code PPC part of library App calls library function, PPE library stub communications to SPEmpeg_encode() MFC Local Store SPU N AUC MFC Local Store SPU N AUC MFC Local Store SPU N AUC MFC Local Store SPU N AUC MFC Local Store SPU N AUC MFC Local Store SPU N AUC MFC Local Store SPU N AUC MFC Local Store SPU N AUC Maintains PowerPC programming model Similiar to IOP but integrated with main processor • Supports shared memory programming • Extremely high on-chip bandwidth • Scales with “conventional” processor SPU code provided in: • Application or middleware libraries • Operating System Services Does not require application rewrite • Specific subsystems targetted • Separate compilations • Leverage ELF • SPUs managed by OS SPE Programming localized in library • SPU Code vectorization required • Private Local Store and DMA model •Scheduling of code / data movement • Code debug and fault isolation complexity Application Specific Accelerators Decoding
  • 48. 48 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Parallel Computational Acceleration Single Source Compiler (PPE and SPE targets) • Auto parallelization ( treat target Cell as an Shared Memory MP ) • Auto SIMD-ization ( SIMD-vectorization ) for PPE VMX and SPE • Compiler management of Local Store as Software managed cache (I&D) Optimization Options • OpenMP-like pragmas • MPI based Microtasking • Streaming languages • Vector.org SIMD intrinsics • Data/Code partitioning • Streaming / pre-specifying code/data use Compiler or Programmer scheduling of DMAs Compiler use of Local store as soft-cache IBM Research Prototype Single Source Compiler • C Frontend • XLC SPE and XLC PPE back-end IBM Research Prototype Parallelizing Compiler • UPC front-end • Fortran front-end • XLC SPE backend 256 GB/sec Coherent Ring MFC Local Store SPU N AU C MFC Local Store SPU N AU C Power Processor (PPE) L2 Cache NCU Local Store SPU MFC N AU C Local Store SPU MFC N AU C N N N N Broadband Engine Chip MFC Local Store SPU N AU C MFC Local Store SPU N AU C MFC Local Store SPU N AU C MFC Local Store SPU N AU C 256 GB/sec Coherent Ring MFC Local Store SPU N AU C MFC Local Store SPU N AU C Power Processor (PPE) L2 Cache NCU Local Store SPU MFC N AU C Local Store SPU MFC N AU C N N N N Broadband Engine Chip MFC Local Store SPU N AU C MFC Local Store SPU N AU C MFC Local Store SPU N AU C MFC Local Store SPU N AU C XDRtm XDRtm I/O I/O Tools and Programmer Driven
  • 49. 49 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Application Source & Libraries PPE object files SPE object files SPESPE SPE SPE Physical SPEs SPESPE SPE SPE Cell AwareOS ( Linux) SPE Virtualization / Scheduling Layer (m->n spe lwp/threads) New SPE lwp/threadsExisting PPE lwp/threads Physical PPE PPE MT1 MT2 Operating System Runtime Strategy Heterogeneous Multi-Threading Model PPE LWP/Threads SPU LWP/Threads SPU DMA EA = PPE Process EA Space OS supports Create/Destroy SPE tasks Atomic Update Primitives used for Mutex SPE Context Fully Managed Context Save/Restore for Debug Virtualization Mode (indirect access) Direct Access Mode (realtime) OS assignment of SPE threads to SPUs Programmer directed using affinity mask
  • 50. 50 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Prototype Cell Extensions to Linux System Call Interface exec Loader File System Framework Device Framework Network Framework Streams Framework SPU Management Framework Privileged Kernel Extensions Firmware / Hypervisor ILP32 Processes LP64 EM Processes Cell Reference System Hardware 32-bit GNU Libs (glibc,etc) 64-bit Linux Kernel 64-bit GNU Libs (glibc) SPUFS FilesystemMisc format bin SPU Object Loader Extension ptrace, large page, madvise, SPE error, fault handling, IIC support, IOS Architecture Specific Code SPU Allocation, Scheduling & Dispatch Extension PPC32 Apps. Cell32 Workloads PPC64 Apps.Cell64 Workloads std. PPC32 elf interp SPE Object Loader Services SPE Management Runtime Library (32-bit) std. PPC64 elf interp Programming Models Offered: RPC, Device Subsystem, Non-Realtime and Realtime Asymmetric Threads -- Single SPU, SPU Groups, Shared Memory SPE Management Runtime Library (64-bit)
  • 51. 51 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Programmer Experience Development Tools Stack System Level Simulator Linux PPC64 with Cell Extensions SPE Management Lib Application Libs Samples Workloads Demos Code Dev Tools Debug Tools Performance Tools Standards: Language extensions ABI Verification Hypervisor Development Environment End-User Experience Execution Environment Cell Prototype Software Environment Miscellaneous Tools
  • 52. 52 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Cell Standards Application Binary Interface Specifications • Defines such things as data types, register usage, calling conventions, and object formats to ensure compatibility of code generators and portability of code. SPE ABI Linux Cell ABI SPE C/C++ Language Extensions • Defines standardized data types, compiler directives, and language intrinsics used to exploit SIMD capabilities in the core. • Data types and Intrinsics styled to be similar to Altivec/VMX. SPE Assembly Language Specification Standards
  • 53. 53 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation System Level Simulator Cell BE – full system simulator • Uni-Cell and multi-Cell simulation • User Interfaces – TCL and GUI • Cycle accurate SPU simulation (pipeline mode) • Pseudo accurate memory and MFC modes • Emitter facility for tracing and viewing simulation events Other simulators • spusim – standalone SPU simulator Execution Environment
  • 54. 54 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Verification Hypervisor (vHype) Seamless Integration of Dual Environments •Low overhead – small footprint •Realtime Resource management •SPE management •Separate policy manager •Pre-emptive partition switching on high priority interrupts Real time Open Source / Proprietary OS Conventional Network OS (Linux) I/O Hosting Hypervisor (Hypervisor - Multi-Partition Resource Allocation / Reservation /Protection) Hardware Networked Apps. Inter-partition Communications Logical Partitioning RTOS/Linux Execution Environment
  • 55. 55 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Linux on Cell All software in STIDC written on Linux OS • Started with Linux 2.4 PPC64 on Cell Simulator SPEs exposed as I/O Devices (function offload model) SPE DMA required pre-pinned memory Inflexible programming model Moved to 2.6.3 • Added heterogenous lwp/thread model – via system call – moved to SPUFS in 2.6.13 SPE thread API created (similar to pthreads library) User mode direct and indirect SPE access models Full pre-emptive SPE context management spe_ptrace() added for gdb support spe_schedule() for thread to physical SPE assignment currently FIFO – run to completion • SPE threads share address space with parent PPE process (through DMA) Demand paging for SPE accesses Shared hardware page table with PPE • SPE Error, Event and Signal handling directed to parent PPE thread • SPE elf objects wrapped into PPE shared objects with extended gld SPE-side mini-loader • madvise() extended for L2 cache and TLB locking/preloading (realtime feature) • All patches for Cell in architecture dependent layer (subtree of PPC64) Publishing Initial CellBE Patches for 2.6.13 (Fall 2005 target) Execution Environment
  • 56. 56 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation SPE Management Library SPEs are exposed as threads • SPE thread model interface is similar to POSIX threads. • SPE thread consists of the local store, register file, program counter, and MFC-DMA queue. • Associated with a single Linux task. • Features include: Threads - create, groups, wait, kill, set affinity, set context Thread Queries - get local store pointer, get problem state pointer, get affinity, get context Groups - create, set group defaults, destroy, memory map/unmap, madvise. Group Queries - get priority, get policy, get threads, get max threads per group, get events. SPE image files - opening and closing SPE Executable • Standalone SPE program managed by a PPE executive. • Executive responsible for loading and executing SPE program. It also services “syscall” requests for I/O (eg, fopen, fwrite, fprintf) and memory requests (eg, mmap, shmat, …). Execution Environment
  • 57. 57 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Optimized Prototype Libraries Audio (resampling) Cryptographic Fast fourier transform Game math Image Large matrix Math Matrix (4x4) Miscellaneous Memory management Multi-precision math Noise & Turbulence Oscillator Parallel programming (MPI like) Shared memory SPU plugin SPU system call Surfaces / curves Synchronization Vector Execution Environment
  • 58. 58 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Code Development Tools GNU based binutils • gas SPE assembler • gld SPE ELF object linker gld extensions for embedding SPE object modules in PPE executables • misc bin utils (ar, nm, ...) targeting SPE modules • hosted on Linux IA32, Linux PowerPC GNU based C/C++ compiler targeting SPE • From STI Partner • retargeted compiler to SPE • Supports common SPE Language Extensions and ABI (ELF/Dwarf2) object output Cell Broadband Engine Optimizing Compiler (IBM Proprietary) • IBM XLC C/C++ for PowerPC (Tobey) • IBM XLC C retargeted to SPE assembler (including vector intrinsics) - highly optimizing • Prototype - XLC Compiler supporting CellBE Programmer Productivity Aids Single Source compilation using OpenMP like pragmas (PPE and SPE object code generated) Auto-Vectorization (auto-SIMD) for SPE code Auto-Parallelization across SPEs UPC Front end – parallelization across SPEs Local Store software managed caching model • Hosted on Linux • Executables to be available on IBM Alphaworks – Fall 2005 Development Environment
  • 59. 59 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Debug Tools CellBE system simulator • Executable availability on AlphaWorks (Fall 2005 target) GNU gdb • ptrace and spe_ptrace enabled • Multi-core Application source level debugger supporting PPE multithreading, SPE multithreading, interacting PPE and SPE threads • Three modes of debugging SPU threads Attach to SPE thread Launch mode – launch a new debug session for each SPE thread Pass-thru mode – follow execution into SPE thread RISCwatch • Low level hardware (JTAG) debugger Development Environment
  • 60. 60 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Prototype Performance Tools pmcount • Tool to access to HW performance counters Performance inspector • Suite of GPL based performance analysis tools extended to support SPE threads tprof – timer based analysis tool ptt – per thread time ai – above idle post – report generator a2n – address to name ctrace • Branch tracing performance monitor (under development) Development Environment
  • 61. 61 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation SPE Performance Tools Static analysis (spexlc_timing) • Annotates assembly source with instruction pipeline state Dynamic analysis (CellBE System Simulator) • Generates statistical data on SPE execution Cycles, instructions, and CPI Single/Dual issue rates Stall statistics Register usage Instruction histogramming Development Environment
  • 62. 62 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Miscellaneous Tools – IDL Compiler Development Environment
  • 63. 63 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Samples / Workloads / Demos Numerous code samples provided to demonstrate system design constructs Complex workloads and demos used to evaluate and demonstrate system performance Terrain Rendering Engine Subdivision Surfaces Physics Simulation Geometry Engine Execution Environment
  • 64. 64 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Subsystem Sample – Geometry Engine OpenGL-like geometry engine •Geometry processing is offloaded to compile-time configurable SPE “vertex shader” •User Queue communication model consisting of 4KB blocks for SPU command requests with command headers in SPE Mailbox FIFO PPE Application G E L I B GE Shader User Queues PPE SPE System Memory Viewer Execution Environment
  • 65. 65 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Terrain Rendering Engine Overview Visualization of Terrain Data Increasingly Important •General availability of high resolution satellite images •Publicly available USGS Digital Elevation Models (DEMs) •Mobile GPS devices (land, sea, air) + wireless networks Inferior Current Solutions •Polygonal Models (low quality, non-Real Time) •Requires CPU + GPU (Graphics Processing Unit) Superior CellBE Solution •Highly Compressed Height Map Models (10x less data) •Requires only one CellBE (No GPU) •High Quality Images (Multi-Sampled Raycast) •Fast (Real Time Animations)
  • 66. 66 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Earthviewer – Polygon Render
  • 67. 67 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Real-Time Terrain Rendering Engine (TRE) Cell Optimized Raycaster Advanced SPUAdvanced SPU shadershader functionfunction Real time rendering with only one Cell processoreal time rendering with only one Cell processor No graphics adapter assistNo graphics adapter assist High definition resolutionHigh definition resolution Ray/Terrain intersection computationRay/Terrain intersection computation Texture FilteringTexture Filtering Normal computationNormal computation Bump map computationBump map computation Diffuse + Ambient lighting modelDiffuse + Ambient lighting model PerlinPerlin Noise based cloudsNoise based clouds Atmosphere computation (haze, sun, halo)Atmosphere computation (haze, sun, halo) Dynamic multi-sampling (4Dynamic multi-sampling (4 –– 16 samples per pixel)16 samples per pixel) Image based input (16 bit height + 16 bit texture)Image based input (16 bit height + 16 bit texture) 47 KB of SPU object code47 KB of SPU object code MJPEG like compression via SPUMJPEG like compression via SPU Performance scales linearly with number of availablePerformance scales linearly with number of available SPUsSPUs Written completely in C withWritten completely in C with intrinsicsintrinsics Client supportClient support –– OpenGL workstation, wireless PDAOpenGL workstation, wireless PDA User inputsUser inputs –– graphical, joystick, GPS/accelerometergraphical, joystick, GPS/accelerometer
  • 68. 68 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation 3D Surface from 25x25 Height Data 0 1000 2000 3000 4000 5000 6000 7.5 KB for just the Surface
  • 69. 69 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Raw 25x25 Height Data 0 1000 2000 3000 4000 5000 6000 1.25 KB for Height Map (6x Compression)
  • 70. 70 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Ray Casting
  • 71. 71 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Height Color Data Layout (Main Memory) © 2005 IBM Corporation71 Quad Word (128 bits)Quad Word (128 bits) H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C
  • 72. 72 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation PPE Data Staging via L2 Main MemoryMain Memory PPEPPE L2L2 SPESPE SPESPESPESPE SPESPE SPESPE SPESPESPESPE SPESPE The L2 has 4 Outstanding Loads + 2 Prefetch
  • 73. 73 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation SPE Data Staging 16 Outstanding Loads per SPE Main MemoryMain Memory PPEPPE L2L2 SPESPE SPESPESPESPE SPESPE SPESPE SPESPESPESPE SPESPE
  • 74. 74 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Height Color Data Layout (Main Memory) © 2005 IBM Corporation74 Quad Word (128 bits) H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C Quad Word (128 bits)
  • 75. 75 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Height Color Data Layout (Local Store) © 2005 IBM Corporation75 SIMD Register (128 bits) H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C H C Shuffle Byte Quad Word (128 bits)
  • 76. 76 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation TRE Image Pipeline Time PPE SPE 0-6 SPE 7 Prep Frame 0Prep Frame 0 Prep Frame 1Prep Frame 1 Render Frame 0Render Frame 0 Render Frame 1Render Frame 1 Encode Frame 0Encode Frame 0 Encode Frame 1Encode Frame 1 Prep Frame 2Prep Frame 2 Prep Frame 3Prep Frame 3 Render Frame 2Render Frame 2
  • 77. 77 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation TRE SPE Render Pipeline MFC Load RC 0 Load RC 0 Load HC/Accm 0Load HC/Accm 0 Load HC/Accm 1Load HC/Accm 1 Store Accm 1 Store Accm 1 Load HC/Accm 0Load HC/Accm 0 Store Accm 0 Store Accm 0 Gen Lists 0Gen Lists 0 Gen Samples 1Gen Samples 1 Gen Lists 0Gen Lists 0Gen Samples 0Gen Samples 0Gen Lists 1Gen Lists 1 Store Accm 0 Store Accm 0 Load RC 0 Load RC 0 Gen Samples 1Gen Samples 1 Load RC 1 Load RC 1 TimeTime SPU
  • 78. 78 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Ray-casting with SIMD Ray 1 Ray 4Ray 3Ray 2 Vector (128 bits) 4 3 2 1
  • 79. 79 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation TRE SPE Ray Kernel Ray IntersectionRay Intersection Multi - Sample AdjustmentMulti - Sample Adjustment Shader Texture Filtering Normal Generation Bump Mapping Lighting Atmosphere Shader Texture Filtering Normal Generation Bump Mapping Lighting Atmosphere Sample AccumulationSample Accumulation
  • 80. 80 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation SPE Local Store Memory Layout Height Color Data (57 KB) Height Color Data (57 KB) Height Color Data (57 KB) Height Color Data (57 KB) Accumulation Data (30 KB) Accumulation Data (30 KB) Accumulation Data (30 KB) Accumulation Data (30 KB) Accum DMA List (15 KB)Accum DMA List (15 KB) Accum DMA List (15 KB)Accum DMA List (15 KB) RCRC RCRC ACAC Stack 4 KBStack 4 KB Code (29 KB) Code (29 KB) HC DMA List (13 KB)HC DMA List (13 KB) HC DMA List (13 KB)HC DMA List (13 KB)
  • 81. 81 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Chip
  • 82. 82 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Sample SPE Simulator Output Performance Cycle count 96187613 Performance Instruction count 113483578 (108381008) Performance CPI 0.85 (0.89) Branch instructions 1255537 Branch taken 826730 Branch not taken 428807 Hint instructions 507711 Hint hit 809520 Single cycle 59886926 ( 62.3%) Dual cycle 24247041 ( 25.2%) Nop cycle 133131 ( 0.1%) Stall due to branch miss 1024965 ( 1.1%) Stall due to prefetch miss 22394 ( 0.0%) Stall due to dependency 9709208 ( 10.1%) Stall due to fp resource conflict 0 ( 0.0%) Stall due to waiting for hint target 1163937 ( 1.2%) Stall due to dp pipeline 0 ( 0.0%) Channel stall cycle 0 ( 0.0%) SPU Initialization cycle 9 ( 0.0%) ----------------------------------------------------------------------- Total cycle 96187611 (100.0%) The number of used registers are 128, the used ratio is 100.00
  • 83. 83 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Austin
  • 84. 84 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Mount Saint Helens
  • 85. 85 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation TRE 720P Performance 2.0 GHz Apple G5 0.6 frames/sec • 40% of cycles spent waiting for Memory 3.2 GHz Cell 30.0 frames/sec • 1% of cycles spent waiting for Memory Cell has 50x advantage
  • 86. 86 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation Summary Cell ushers in a new era of leading edge processors optimized for digital media and entertainment Desire for realism is driving a convergence between supercomputing and entertainment New levels of performance and power efficiency beyond what is achieved by PC processors Responsiveness to the human user and the network are key drivers for Cell Cell will enable entirely new classes of applications, even beyond those we contemplate today
  • 87. 87 IBM Systems and Technology Group STI Technology © 2005 IBM Corporation (c) Copyright International Business Machines Corporation 2005. All Rights Reserved. Printed in the United Sates September 2005. The following are trademarks of International Business Machines Corporation in the United States, or other countries, or both. IBM IBM Logo Power Architecture Other company, product and service names may be trademarks or service marks of others. All information contained in this document is subject to change without notice. The products described in this document are NOT intended for use in applications such as implantation, life support, or other hazardous uses where malfunction could result in death, bodily injury, or catastrophic property damage. The information contained in this document does not affect or change IBM product specifications or warranties. Nothing in this document shall operate as an express or implied license or indemnity under the intellectual property rights of IBM or third parties. All information contained in this document was obtained in specific environments, and is presented as an illustration. The results obtained in other operating environments may vary. While the information contained herein is believed to be accurate, such information is preliminary, and should not be relied upon for accuracy or completeness, and no representations or warranties of accuracy or completeness are made. THE INFORMATION CONTAINED IN THIS DOCUMENT IS PROVIDED ON AN "AS IS" BASIS. In no event will IBM be liable for damages arising directly or indirectly from any use of the information contained in this document. IBM Microelectronics Division The IBM home page is http://www.ibm.com 1580 Route 52, Bldg. 504 The IBM Microelectronics Division home page is Hopewell Junction, NY 12533-6351 http://www.chips.ibm.com