FPGA/Reconfigurable computing (HPRC)

May 10, 2014 R.Innocente 1
Reconfigurable ComputingReconfigurable Computing
Roberto Innocente
inno@sissa.it

Flexibility
ASIC
Application
Specific
Integrated Circuit
Very inflexible,designed
to solve just 1 problem.
Energy, space and time
efficient
GPP
General
Purpose
Processor
Very flexible,
can solve any problem.
Energy, space and time
inefficient
?
Reconfigurable
Hardware
Flexible,
But enough energy, time
and space efficient
+-

History

Gerald Estrin/1
is credited with the idea, in the
'60, of the first reconfigurable
(F+V) FIX+Variable computer
Gerald Estrin. ACM 1960. Organization of computer
systems: the fixed plus variable structure computer.

Gerald Estrin/2
He envisioned that important gains in performance could be achieved when
many computations are executed on appropriate problem oriented configurations.
F+V is made of :
- high speed general computer(the F part) : initially an ibm7090
- various size high speed special structures (the V part) problem specific: trigonometric functions, logarithms,
exponential, n-th powers, complex arithmetic, …
V is made of a 36 module positions motherboard which can undergo :
- Function reconfiguration: physically changing some modules
- Routing reconfiguration : changing part of the back wiring
The Rammig machine (1977) : investigation of a reconfigurable machine with no manual or mechanical
intervention

Today reconfigurable hardware
Is born out of the will to replace different logic
IC(Integrated Circuits), and successively to rapidly
prototype large ASICs(Application Specific ICs) or
implement SoCs (Sytem On Chip).
In the following slides readers are supposed to be involved
in scientific computing and not EE engineers.

Basic digital circuits
AND INVERTER
Shift RegD Type FFMUX
Usually 0=0V, 1=some positive voltage
OR

SSI 74xx IC

PLD
Inconvenience of standard discrete logic circuits :
- 14 pin packages of 4/6 logic functions
- often you had to traverse the PCB to find a free OR or inverter
- if you needed only a few, you had in any case to put an IC with 4/6
Therefore came the idea of PLD (Programmable Logic Device) :
- SPLD (Simple : PAL/PLA)
- CPLD (Complex)
In which a simple interconnection network could be configured melting some internal
fuses (fuse technology) to implement combinatorial logic.

disjunctive normal form
(aka Sum of products )
Each boolean function of some boolean variables can be
represented as a sum of minterms (product of all variables
or their complement) .
With 3 boolean vars : a,b,c
are 2 of the 23 = 8 minterms
f (a ,b , c)=a ̄b c+̄a b ̄c
ābc,̄ab̄c

PLA (Programmable Logic Array)
f1= p1+ p2 + p3=x1x2 + x1 ̄x3+ ̄x1 ̄x2 x3+ x1 x3

FPGA
Also CPLDs showed their limits, therefore in 1985/1990
Xilinx introduced a more flexible design , the
FPGA (Field Programmable Gate Array)
In which the interconnection network is much
more flexible and on which also sequential
circuits can be easily mapped.

FPGA idea
1985 Xilinx – Ross Freeman (inventor of
FPGA): “What if we could develop the
equivalent of a circuit board full of
standard logic parts (like TTL and PAL
devices) on a single high density
programmable logic chip ?”
- post fabrication programmability by end
users
- fabless semiconductor company

Today

FPGA market
Dominated by 2 players :
- Altera
- Xilinx
From 67% of 2010, today they share together 90% of the market
(4.5 billion usd revenues in 2012)
From sourcetech411(2010)

An important question: are FPGAs green ?
Virtex-7 2000T (one of the
top FPGAs) :
~ 20 W
Xilinx showed 3600 copies of its 8
bit processor nanoblaze running on
Virtex-7, consuming 20 W
CPU : ~ 100 W
Core i7-4770K Haswell (22 nm) 3.5 GHz@ 4 Cores 84 W
Core i7-3930K Sandybridge-E (32 nm) 3.2 GHz @6Cores 130 W
Xeon E7458 Dunnington (45 nm) 2.4 GHz 90 W
Xeon E7460 Dunnington (45 nm) 2.66 GHz 130 W
GPU : ~ 220 W
Nvidia Tesla M2090 225 W
Nvidia Tesla K20X 235 W
This is a partial answer. We need to be able to estimate FPGA
performance to give a more useful index.

FPGA architecture
From RF and
Wireless World
Sea of gates : logic blocks are like islands in a sea of interconnections

Virtex family
1998 Virtex 250nm 100mhz 25k-60k cells
2000 Virtex-E 180nm 300mhz 1k-70kcells
2000 Virtex II 150nm to168 mult420mhzupto 93k 4-luts
2005 Virtex-4 90nm 500mhz upto 200k cells
2007 Virtex-5 65nm 550mhz up to 330k cells
Virtex-6 40nm 288-2k DSP to 500k 6-luts
2010 Virtex-7 28nm ~500mhz upto 2000k cells
2014 Virtex-US 20 nm upto 4400k cells
From L Zhuo
Up to ~ 7 billion transistor
Intel 2014 15-core Xeon IvyBridge-EX~ 4.3 billion transistor
Nvidia 2012 GK110 Kepler ~ 7 billion transistor

FPGA/CPU evolution

Virtex-7 is not monolithic
2.5 D technology : 4 FPGA tiles with silicon interposer that provides 10k
Interconeections between layers

Enabling technologies

Programming technology/1
Antifuse SRAM
OTP(One time programmable)
Disordered except at very low range
Pass transistor in switch block

Programming technology/2
Antifuse
-pros:
cheap, small
-cons:
requires special
processing, One time
programming
SRAM
-pros:
can be deployed with standard
semiconductor process, can be
easily reprogrammed
-cons:
large area required(6
transistors)

Confware
The configuration of an FPGA ( that becomes compiled to a
stream of bits) is not hardware, nor software.
Someone invented the neologism
confware
The configuration of a reconfigurable hardware.

How you configure an FPGA ?
SRAM cells as a long shift register : loaded serially clocking in the confware
Virtex 7 2000T = 440 Mbits of SRAM cells
(simplified : large fpgas can also parallel load the confware)

Logic Blocks/Logic Cells

Fine/coarse grain logic blocks
From :
- a single transistor
(Crosspoint : went in
bankrupcy)
- a logic gate
To :
- a complete processor (FPNA: field
programmable node arrays)
NB. FPNA is also field programmable neural array

Homogeneous :
- Logic Cells: 4 input LUT(LookUp Table) + FlipFlop
Heterogeneous(modern development) :
- Logic cells
- DSP (Digital Signal Processing)
- Memory blocks
- I/O blocks
The heterogenous architecture is prevalent now. The blocks are configured by SRAM bits usually
loaded trough serial ports as already pointed out.
CLB(Configurable Logic Blocks)
Necessary differentiation to allow
things like multiplication/addition
to be mapped in an efficient way.

Standard Logic Cell
4 input LUT
D type FlipFlop
16 bits of SRAM for conf 1 bit SRAM conf
2:1 Mux

standard LUT (Look Up Table)
0 0000 0
1 0001 1
2 0010 0
3 0011 0
4 0100 1
5 0101 0
6 0110 1
7 0111 1
.. .. ..
Dec Bin Out
- 16 x 1 memory
- any boolean function of 4
inputs :
Bit 0
Bit 1
Bit 2
Bit 3
f = ̄x3 ̄x2 ̄x1 x0+ ̄x3 x2 ̄x1 ̄x0+ ̄x3 x2 x1 ̄x0+ ̄x3 x2 x1 x0
NB. LUT rhymes
with nut

Uses of Logic Cell
2^4 = 16 x 1 bit memory Any boolean function of 4
inputs
4:1 multiplexer

Virtex-7 Logic Block basics

Virtex-7 Logic slice
From Xilinx
4 x 32=128 bit shift reg

Virtex7 CLB slice
- 6-input LUT
- 2 5-input LUTs with same inputs
- 2 arbitrary boolean function on 3-input and 2-input or less

Altera ALM

Interconnection network

Interconnection network
Hierarchical routing Island type routing(predominant)
Interconnection network can consume
80% of the area of an FPGA !
Nearest neighbours

Programmable switch

SRAM routing: coarse/fine grain
5 bit SRAM 1 bit SRAM

Details of island type routing

Disjoint/Wilton switch blocks
Disjoint : wire can only go out on
wire of same number, creates routing
domains
Wilton : can change domain in at
least one directions

Channel segments distribution

Columnar architecture
7 series Xilinx fpga
Columnar architecture

DSP blocks &
floating point

FPGAs floating point in 1994
B. Fagin and C. Renard. Field Programmable Gate Arrays and
Floating Point Arithmetic. IEEE Transactions on VLSI Systems,
2(3), September 1994.
Fagin & Renard report that you can implement
floating point operators but it is impractical : no
FPGA in existence could contain a single
multiplier circuit !!

FPGA fp in 1995
Shirazi & al. On the same line of Fagin & Renard propose
2 custom fp formats 16 and 18 bits total:
they provide for them add,sub, mul, div operators
N. Shirazi, A. Walters, and P. Athanas. Quantitative Analysis of
Floating Point Arithmetic on FPGA Based Custom Computing
Machines. In Proceedings of the IEEE Symposium on FPGAs for
Custom Computing Machines, April 1995.

FPGA fp in 2002
Belanovic & Leeser present a library of variable width
parameterized floating point operators (superset of the
ieee formats)
A Library of Parameterized Floating-point Modules
and Their Use
Pavle Belanovic and Miriam Leeser, 2002

What allowed the breakthrough ?
The addition, by major vendors, of hardware multipliers (called
DSP blocks) on their FPGA from 2000 on :
- 1st Xilinx on Virtex II
- soon after Altera on Stratix
This started in the last decade also the interest of HPC community :
Cray XD1, Silicon RASC, Convey HC1
HPRC = High Performance Reconfigurable Computing

FPGA MAC operation

Virtex-7 DSP48 high level
From Xilinx
1 bit 2 bit

DSP48E1 details

Altera Stratix V DSP block
4 (*) + 3(+) = 7 flop

Data Flow
Graphs (DFG)

Data flow
A representation of a program as a DG(Directed Graph) in
which the nodes are the operations and the edges represent
the data dependencies from one operation to the next

Control flow/Data Flow
dis2=b**2-4*a*c
If dis2 < 0 complex!
dis=sqrt(dis2)
u1=-b/(2*a)
u2=dis/(2*a)
x1=u1+u2
x2=u1-u2
x=
−b
2a
±
√b2
−4ac
2a

A scalar product
Fortran :
acc=0.0
do i=1,4
acc=acc+a(i)*b(i)
enddo
C :
acc=0.0;
for(i=0;i<4;i++){
acc=acc+a[i]*b[i];
}

Time/Space tradeoffs

Systolic array matrix mult
A(n,n) x B(n,n) requires :
2n-1 steps for the last elements to enter the array
n-1 steps to compute the last c(n,n)
n steps to move the result out = 4n-2 steps

Codesign
The implementation of algorithms on FPGAs requires a mix
of hw and sw design :
Codesign = hw design + sw design

How to program FPGAs?
Mainly with an HDL (Hardware Description Language):
- Verilog(intially developed by Gateway Design Automation, now a std)
- VHDL (out of a standard committee)
But OpenCL, ImpulseC, SystemC, C, Handel-C translators .. are also available.Is this a good idea ?
The problem is that those languages are not thought for describing hardware and the translation finish
up usually with a FSM(finite state machine) with 1 state for every statement and then the FSM
machine moves along the states .
This is not the way someone skilled would program the FPGA.
Next state
logic
State
register
Output
Logic
input
clk
D Q
Out
FSM finite state
machine

FPGA will win
For many years FPGAs were just prototyping vehicles for
ASICs
– Now they are replacing many ASICS & ASSPs
– Watch for the same Trojan effect with FPGAs in HPC

FPGA lingo

Core
Core in FPGA lingo is a function ready to be instantiated
into your design as a “black box”. It can be suppliad as
HDL or schematic.
It supports design re-use.

Soft/hard cores
On FPGAs functional modules can be implemented :
- using std FPGA resources(logic blocks, DSPs, memory
blocks) : softcores
- as an ASIC on the FPGA : hardcores
When the manufacturer puts a processor as an hardcore on
the FPGA then it sells this as a SoC (Sytem On Chip) :
Dual ARM on Zync-7000 chip, PowerPC on Altera FPGA

IP/open cores
The soft attribute is implied.
Hardware designs in an HDL(eventually using vendor libraries):
- opensource cores : http://opencores.org/
OpenRISC 1000 architecture from the OpenCores community,
the Lattice Semiconductor LM32, the LEON3 from Aeroflex
Gaisler and the OpenSPARC family from Oracle
- proprietary : IP(Intellectual Property) cores
Floating point operators, fft, matrix computations

Commercial offers

Picocomputing
SC6 1U Upto 16 FPGA SC6 4U upto 48
EX-600EX-800
From
PICOCOMPUTING

Bittware Terabox
16 altera stratix-V
From Bittware

DINIGROUP Cluster of 4 Virtex7
From
DINIGROUP

Dinigroup Cluster 40 Kintex-7
From DINIGROUP

Maxeler MPC-X
Daresbury Lab UK :
The dataflow supercomputer
will feature Maxeler developed
MPC-X nodes capable of an
equivalent 8.52TFLOPs per
1U and 8.97 GFLOPs/Watt.

Convey HC-2 , HC-2ex

Cray XT5h
“Cray introduces an hybrid supercomputer
that
can integrate multiple processor architectures
into a single system and accelerate high
performance computing (HPC) workflows.
The Cray XT5h delivers higher sustained
performance, by applying alternative
processor architectures across selected
applications within an HPC workflow. The
Cray XT5h supports a
variety of processor technologies, including
scalar processors based on AMD OpteronTM
dual and quad-core technologies, vector
processors, and FPGA accelerators.”

CHREC
Center for High Performance
Reconfigurable Computing
UF/BYU/GWU/VTECH

CHREC Novo-G 384 FPGAs
“Novo-G is the most powerful
reconfigurable supercomputer in
the known world. This unique
machine features 192 top-end,
40nm FPGAs (Altera Stratix-IV
E530) and 192 top-end, 65nm
FPGAs (Stratix-III E260). “
http://www.chrec.org/
(pronounce it as shreck)

BLAST like Smith-Waterman computes local alignment of 2
sequences :
- Novo-BLAST Novo-G/CHREC implementation : faster, same
sensitivity
IPC(Isotope Pattern Calculator) of Protein Identification Algorithm :
- speed up 52-366 on single fpga, 1259 on 4 fpgas, 3340 on a node(16
fpgas)
CHREC/2

References for
Applications

Linear Algebra for RC
Juan Gonzalez and
Rafael C. Núñez
LAPACKrc: Fast linear
algebra kernels/solvers
for FPGA
accelerators(JP 2009)
DOD funded

DCT, FFT on FPGAs
Digital Signal Processing with Field Programmable Gate
Arrays ,
3d edition(2007)
U.Mayer Baese, Springer Verlag

MD on FPGA
There are many papers about porting Molecular Dynamics algorithms on
FPGAs with substantial positive conclusions about experiments on 1-2
FPGAs. But in the last years there is an embarassing comparison with
ANTON (Shaw et al.).
We cant forget that ANTON is a really huge machine consuming over 100
KW !!!!
And is made out of 512 dedicated ASICs at 1ghz!
The comparison with some FPGAs consuming 40/60 W is improper.
FPGA-Accelerated Molecular Dynamics(2013) M. A. Khan,M. Chiu, M. C. Herbordt

Neural networks on FPGAs
Editors : Omondi , Rajakapse (2006)
FPGA implementation of neural networks
ANN(Artificial Neural Network) in integer arithmetic
performs 40x better than on GPP (old FPGA, 3
generation old)

Altera Arria 10

Arria10

Arria 10 variable precision DSP block
Altera
A
B
C
D
A+C*D = 2 flop

Arria10 estimated sp fp performance
- 2 flops per cycle
- 1688 fp single precision DSP (GX660)
1688*2 = 3376 flops per cycle
3376 * 0.5 ghz ~ 1.7 Teraflops in single precision

Hard single prec FP on FPGA ?!?
For people that can live with single precision this seems a very
attractive new feature.
But many think that it is too much a waste of generic resources
and claim that what was missing were simpler blocks !

Back of the envelope
performance estimation

Back of the envelope performance estimation
Given number of
- LUTs
- FFs
- DSPs
offered by an FPGA,
and utilization of resources
by operators, estimate the
max number of operators
that can be implemented on
the FPGA
Today FPGA clocks are ~500Mhz=0.5GHz
(unavoidable price for flexibility)
2000 flops per cycle = 1 Teraflops

Xilinx Virtex-7 family
Virtex-7 slices : 4 x 6-input LUTs, 8 FFs
Virtex-7 DSPs : 48 bits pre-adder, 25x18 multiplier, 48 bits accumulator
Virtex LUT ~ 1.6 standard LUT

Custom precision 17/24 bits floating
dsp lut+f lut f # tot dsp tot lut tot f
* 2 103 90 112 1080 2160 208440 232200
1 113 97 104 0 0 0 0
0 377 336 376 0 0 0
0 0 0 0 0 0 0
0 0 0
+ 0 369 301 393 1510 0 1011700 1150620
0 0 0 0 0 0 0 0
Tot 2590 2160 1220140 1382820
Virtex-7 V2000T available resources
slices LUT x FF x dsp 6 input ff
slice slice LUT
305400 4 8 2160 1221600 2443200
1.6
standard LUTs 1954560

IEEE single precision – 32 bits
* 3 120 103 105 700 2100 156100 157500
2 160 128 160 0 0 0 0
1 331 283 331 0 0 0
0 665 629 669 0 0 0
0 0 0
+ 2 293 225 327 25 50 12950 15500
0 500 407 541 1160 0 1052120 1207560
Tot 1885 2150 1221170 1380560
slice slice LUT
305400 4 8 2160 1221600 2443200
1.6

IEEE double precision – 64 bits
* 11 325 279 421 196 2156 118384 146216
10 371 299 456 0 0 0 0
9 439 356 510 0 0 0
0 2361 2317 2418 0 0 0
0 0 0
+ 3 895 705 945 1 3 1600 1840
0 989 794 1029 617 0 1100111 1245106
Tot 814 2159 1220095 1393162
slice slice LUT
305400 4 8 2160 1221600 2443200
1.6

Virtex UltraScale XCVU440 20nm -sampling out
IEEE double precision – 64 bits
* 11 325 279 421 261 2871 157644 194706
10 371 299 456 0 0 0 0
9 439 356 510 0 0 0
0 2361 2317 2418 0 0 0
0 0 0
+ 3 895 705 945 3 9 4800 5520
0 989 794 1029 1321 0 2355343 2665778
Tot 1585 2880 2517787 2866004
Virtex Ultra Scale - available resources
slice slice LUT
314820 8 16 2880 2518560 5037120
1.6

Relative power dissipation/1
TDP/peak nominal double fp performance :
Intel Q6600 2.4ghz 105W/ 38 gflops = 2763mW/gflops
Intel Haswell i7-4770K 3.5ghz 84W/ 112 gflops = 750mW/gflops
Intel IvyBridge 3770K 3.5ghz 77W/ 112 gflops = 687mW/gflops
Nvidia Tesla M2090 225W/ 666 gflops = 337mW/gflops
Nvidia Tesla K20X 235W/1310gflops = 179mW/gflops
Xilinx Virtex-US 20W/ 800gflops = 25mW/gflops
Ro w 1
0
FPGA computing = green computing
}
} ~10x
~30x

Relative power dissipation/2
Intel 2.4 ghz q6600
intel 4770k
intel i7-3770k
tesla m2090
tesla k20x
virtex7
0 500 1000 1500 2000 2500 3000
mW / Gflops
mW

Gflops per Watt
peak nominal double fp performance/TDP :
Intel Q6600 2.4ghz 38 gflops/105 W = 0.36 gflops/W
Intel Haswell i7-4770K 3.5ghz 112 gflops/84 W = 1.33 gflops/W
Intel IvyBridge 3770K 3.5ghz 112 gflops/77 W = 1.45 gflops/W
Nvidia Tesla M2090 666 gflops/225 W = 2.96 gflops/W
Nvidia Tesla K20X 1310 gflops/235 W = 5.57 gflops/W
Xilinx Virtex-US 800 gflops/20 W = 40 gflops/W
Ro w 1
0
FPGA computing = green computing
}
} ~10x
~30x

Top green500 list
green500_ranktotal_power Year name Total CoresName ManufacturerCountry
1 28 4,503 2013 2720TSUBAME-KFC NEC Japan
2 53 3,632 2013 5120Wilkes Dell United Kingdom
3 79 3,518 2013 4864HA-PACS TCA Cray Inc. Japan
4 1,754 3,186 2012 115984 Cray Inc. Switzerland
5 81 3,131 2013 5720romeo Bull SA France
6 923 3,069 2013 74358TSUBAME 2.5 NEC/HP Japan
7 54 2,702 2013 3080 IBM United States
8 270 2,629 2013 15840 IBM Germany
10 71 2,359 2010 4620CSIRO GPU Cluster Xenon SystemsAustralia
11 179 2,351 2012 38400SANAM Saudi Arabia
13 82 2,299 2012 16384Cetus IBM United States
14 82 2,299 2012 16384 IBM Poland
16 82 2,299 2012 16384Vesta IBM United States
18 237 2,243 2013 10920HPCC Hewlett-PackardUnited States
Mflops/Watt
LX 1U-4GPU/104Re-1G Cluster, Intel Xeon E5-2620v2 6C 2.100GHz, Infiniband FDR, NVIDIA K20x
Dell T620 Cluster, Intel Xeon E5-2630v2 6C 2.600GHz, Infiniband FDR, NVIDIA K20
Cray 3623G4-SM Cluster, Intel Xeon E5-2680v2 10C 2.800GHz, Infiniband QDR, NVIDIA K20x
Cray XC30, Xeon E5-2670 8C 2.600GHz, Aries interconnect , NVIDIA K20xPiz Daint
Bull R421-E3 Cluster, Intel Xeon E5-2650v2 8C 2.600GHz, Infiniband FDR, NVIDIA K20x
Cluster Platform SL390s G7, Xeon X5670 6C 2.930GHz, Infiniband QDR, NVIDIA K20x
iDataPlex DX360M4, Intel Xeon E5-2650v2 8C 2.600GHz, Infiniband FDR14, NVIDIA K20x
iDataPlex DX360M4, Intel Xeon E5-2680v2 10C 2.800GHz, Infiniband, NVIDIA K20x
iDataPlex DX360M4, Intel Xeon E5-2680v2 10C 2.800GHz, Infiniband, NVIDIA K20x
Nitro G16 3GPU, Xeon E5-2650 8C 2.000GHz, Infiniband FDR, Nvidia K20m
Adtech, ASUS ESC4000/FDR G2, Xeon E5-2650 8C 2.000GHz, Infiniband FDR, AMD FirePro S10000Adtech
BlueGene/Q, Power BQC 16C 1.60 GHz, Custom
BlueGene/Q, Power BQC 16C 1.600GHz, Custom Interconnect
BlueGene/Q, Power BQC 16C 1.60GHz, Custom
BlueGene/Q, Power BQC 16C 1.60GHz, Custom
Cluster Platform SL250s Gen8, Xeon E5-2665 8C 2.400GHz, Infiniband FDR, Nvidia K20m

Power/Energy efficiency

Power Dissipation
PT =k C V
2
f +Ps
Ed=
1
2
C V 2
A chip is made of millions of CMOS FETs. When
input switches, you need to charge the small
capacitance :
f times a second gives, together with some
constant static dissipation :
Anyway increasing a lot the frequency, the
chip becomes unstable unless you increase
also the voltage(leakage). Therefore there is
in fact a superlinear behaviour vs f:

Dennard scaling(1974)
1
S
S3
S2
= 2x more
transistors
S = 1.4x lower
capacitance
Scale Vdd by S =>
S2
= 2x lower
energy
S2
S = 1.4x faster
transistors
Performance scales as S3
= 2.8 while power density stays constant across
generations

Fred Pollack(Intel) famous graph(1999)
Power density increases !!!
In 2004/2005 we hit the power wall => stop frequency
increases
“New
microarchitecture
challenges in the
coming generations
of CMOS process
technology”
F.Pollack

End of Dennard scaling
1
S
S3
S2
= 2x more
transistors
S = 1.4x lower
capacitance
S2
S = 1.4x faster
transistors
In submicron
technology rigidity
in voltage scaling.
Power increases by
S2
= 2

MOS subthreshold current
Scaling down geometry you scale down drain voltage to
avoid high electric fields and to decrease energy required to
switch. You have to scale down also the threshold voltage
to sustain the 30% decrease of gate delay. The small voltage
swing that remains is not able to completely turn off the
transistor. Subthreshold leakage that was ignored in the
past can on modern VLSI chips consume up to ½ of the
total power.

Subthreshold leakage

VT
design tradeoff
VGS
log IDS
- Low VT
for high ON current :
- High VT
for low OFF current
Phenomenology :
60-200 mV of VGS
swing decreases IDS
by
one order of magnitude. Today 0.5-0.2
VT
doesn't allow the needed swing of VGS
to
shutoff the transistor.
I Dsat ∝(V DD−VT )2
Low VT
=> high IDS
good for ON condition
High VT
=> low leakage
good for OFF condition

Multicore scaling
65 nm 45 nm 32 nm
4-core 8-core 16-core
Every generation 2x cores, at same or slightly
increasing frequency.

Multicore scaling at constant frequency
1
S
S2
S2
= 2x more
transistors
S = 1.4x lower
capacitance
} S = 1.4x lower
utilization
We hit the utilization wall => dark silicon

End of multicore scaling
65 nm 32 nm
4 cores 8 cores
Every generation 1.4x cores, at same or
slightly increasing frequency.
Dark or dim silicon
(“uncore”)
45 nm
5.7 cores

Dark silicon and the end of multicore scaling
Doug Burger (Microsoft) at HiPEAC 2013 :
- till 2004: each semiconductor generation gave transistors
smaller, faster and that consume less
- from 2004 to now: we still got smaller transistors, but we
could not run them faster (power wall)
- in the future : we will still get smaller transistors but we
will not be able to use all of them together(dark silicon) or
at max speed.

Scaling the utilization wall
G.Venkatesh ASPLOS 10 :
“while the area budget continues to increase exponentially, the power budget has
become a first-order design constraint in current processors. In this regime, utilizing
transistors to design specialized cores that optimize energy-per-computation
becomes an effective approach to improve the system performance.
”The Utilization Wall : With each successive process generation, the percentage of a
chip that can switch at full frequency drops exponentially due to power constraints.
[Venkatesh, ASPLOS ‘10]
Single chip heterogeneous computer (E.Chung)
Greater energy efficiency combining GPP with unconventional cores (U-cores) :
GPU,FPGA,DSP,ASICs ..

3D FinFET promise
Below 20nm the roadmap is
to use 3D FinFETs :
- Faster : +37%
- Dynamic Power: -50%
- Static Power: -90%
KAIST demonstrated a 3nm
FinFET in lab

The trouble with multicore
A famous article of David Patterson (of “Computer architecture:
a quantitative approach” fame) on IEEE Spectrum, 2010 :
“Chipmakers are busy designing microprocessors that most
programmers can’t program”
“... the semiconductor industry threw the equivalent of a Hail
Mary pass when it switched from making microprocessors run
faster to putting more of them on a chip - doing so without any
clear notion of how such devices would in general be
programmed. The hope is that someone will be able to figure out
how to do that, but at the moment, the ball is still in the air.”

Verilog

Using Verilog
You write a functional specification (usually) splitted in
modules that documents the exact behaviour of the system.
Logic
Synthesis
Place &
Route
HDL
(Verilog)
FPGA
ASIC
Functional
design
Physical
design
Gate
netlist
Simulated annealing
used here !
NB. place and route of a large design can take 1
day of a fast CPU !!

Verilog/1
Basic module :
// comments in this way
module name(input x0,x1,input [3:0]y, output out);
// x0,x1 are wires, y is a 4 wires bus
// out is an output wire
// combinational logic use assign
wire x0,x1, [3:0]y, out
endmodule

Verilog/2
Combinatorial circuit :
// performs not a b c + a not b not c
module dummy(input a,b,c, output y,z);
wire a,b,c,y;
assign y = ~a & b & c | a & ~b & ~c;
assign z = ~c;
endmodule
This is not C !
a,b,c,y,z are wires and y,z change whenever
a or b or c change. To avoid this drama for complex circuits
we use synchronous logic
(everything is stepped in docking stations = Flip flops)

Verilog/3

Verilog/4
A sequential circuit :
// a flip flop described in verilog
module ff(input d, clk, output q, qbar);
wire d, clk;
reg q, qbar;
always @(posedge clk)
begin
q <= d;
qbar <= ~d;
end
endmodule
At a raising edge of the wire clk copy the signal to q and
the inverse of d to qbar

Verilog/5

Verilog/6
A more complicate sequential circuit :
// in verilog FF with clear/reset
module ff(input d, clk,clr, output q, qbar);
wire d, clk;
reg q, qbar;
always @(posedge clk, posedge clr)
if (clr)
q <= 0;
else
begin
q <= d;
end
endmodule
At a raising edge of the wire clr set q=0, at the raising edge
of clk copy the signal to q and the inverse of d to qbar

Verilog/7

BORPH : Berkeley Operating system for
ReProgrammable Hardware
PETALINUX : Xilinx linux for Zynq et al.

- Idea of HW unix process :
has pid, can be killed like a
normal unix process, but in
fact is an HW instance on
FPGA
- ioreg Virtual File System
interface
Borph : Berkeley Operating System

Xilinx Petalinux
The PetaLinux Software Development Kit (SDK) is a
development tool that contains everything necessary to build,
develop, test and deploy Embedded Linux systems on : Zync-
7000, Zedboard, Kintex-7 boards.
PetaLinux consists of : pre-configured binary bootable images,
fully customizable Linux for the Xilinx device, and PetaLinux
SDK which includes tools and utilities to automate complex
tasks across configuration, build, and deployment.
PetaLinux is offered under two separate licenses :
No charge Evaluation license or Commercial licenses

END

FPGA/Reconfigurable computing (HPRC)

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to FPGA/Reconfigurable computing (HPRC) (20)

More from rinnocente (16)

Recently uploaded (20)

FPGA/Reconfigurable computing (HPRC)