Intel’S Larrabee
- 2. IntroductionIt is a multicore general purpose graphics processor unit (GPGPU), combines the functions of multi core CPU & GPU.
- 3. Larrabee is based on Intel’s x86 architecture.Architectural convergence
- 4. FeaturesTexture filtering, rasterization, depth testing and alpha blending entirely in softwareImplement binned renderer to increase parallelism Reduced memory BandwidthParallel processing on image processing, physical simulation, medical & financial analysis.DDR5 RAM supportEach core can execute 32Gigaflops/s with 1GHz clock, results several teraflops/s speed
- 5. Differences with CPUOut of order executionVector processing unit supports 16-single precision floating point numbers at a timeTexture sampling units – trilinear /anisotropic filtering & texture decompression1024-bit ring bus between coresCache control instructions4-way multithreading
- 6. Difference with GPUx86 instruction set with Larrabee-specific extensions cache coherency across all its coresz-buffering, clipping, and blending without using graphics hardware
- 7. Larrabee – Block Diagram
- 8. ArchitectureCores communicate on a 1024-bit wide ring bus - Fast access to memory, I/O interfaces and fixed function blocks - Fast access for cache coherencyL2 cache is partitioned among the cores - Provides high aggregate bandwidth - Allows data replication & sharing Optimized for highly parallel workload using vector processorIn-order CPU Core Separate scalar & vector units with separate registers
- 11. Fast access from 64k L1 cache
- 12. Direct connection to eachcore’s subset of the 256k L2 cachePrefetch instructions load L1and L2 caches
- 13. Vector Unit Vector complete instruction set – Scatter/gather for vector load/store – Mask registers select lanes to write, which allows data-parallel flow control – Masks also support data compaction Vector instructions support – Full speed when data in L1 cache – Fused multiply add (three arguments) – Int32, Float32 and Float64 data – Can read 8-bit unorm, 8-bit uint, 16 bit sine, 16 bit float data & convert it into 32 bit floats/ integers.
- 14. Fixed Function LogicMicro codes in place of fixed function logic for post shader alpha blending, rasterizationand interpolation.Includes fixed function texture filter logicVirtual memory for textures
- 15. Larrabee’s Binning RendererBinning pipeline– Reduces synchronization– Front end processes vertex & geometry shading– Back end processes pixel shading, stencil testing, blending– Bin FIFO between them• Multi-tasking by cores– Each orange box is a core– Cores run independently– Other cores can run othertasks, e.g. physics
- 16. Back-end Rendering a Tile• Orange boxes represent work on separate threads• Three work threads do Z, pixel shader, and blending• Setup thread reads from bins and does pre-processing• Combines task parallel, data parallel, and sequential
- 17. Pipeline can be changedParts can move between front end & back end – Vertex shading, tesselation, rasterization, etc. – Allows balancing computation vs. bandwidthNew features – Transparency, shadowing, ray tracing etc. – Each of these need irregular data structures– Also helps to be able to “repack” the data
- 18. TransparencyTransparency with & without pre-resolve effects
- 19. Examples of using TasksApplications – Scene traversal and culling – Procedural geometry synthesis – Physics contact group solve – Data parallel strand groups – Distribute across threads/cores using task system – Exploit core resources with SIMDLarrabee can submit work to itself! – Tasks can spawn other tasks – Exposed in Larrabee Native programming interface(c/c++compiler)
- 21. Scalability Studies Based on memory Bandwidth & texture filtering speedPerformance Breakdowns
- 22. Binning & Bandwidth StudiesBandwidthImmediate mode use more Bandwidth -2.4 to 7 times for F.E.A.R -1.5 to2.6 times more for Gears of War -1.6 to 1.8 times more for Half Life 2 Episode 2.
- 24. Conclusion The Larrabee architecture opens the rich set of opportunities for both graphics rendering and throughput computing and is the appropriate platform for convergence of GPU & CPU
- 25. ReferenceIEEE Digital Library- Larrabee: a many- core x86 architecture for visual computing: - Larry Seiler, Doug Carmean, Toni Juan of Intel Corporation, Jeremy Sugerman & Peter Hanrahan – Stanford UniversityIEEE spectrum January 2008ACM transactions on graphics-Article 18www.intel.comwww.wikipedia.com
- 26. Questions
- 27. Thank You