commercial in confidence. copyright ©2005, nallatech.1 is it time for von neumann and harvard to...

Commercial In Confidence. Copyright ©2005, Nallatech.1

Is it time for Von Neumann and Harvard

to Retire?

Presented by :Allan Cantle – CEO

www.nallatech.com


Agenda

» History & Commercial Realities of FPGA Computing» Thoughts on the future possibilities for FPGA

Computing» FPGA Coprocessor vs FPGA main processor» Optimizing Spatial & Temporal Demands of Computing

Problems» Homogeneous vs Heterogeneous vs Polymorphic Computing» Coarse Grain Vs Fine Grain Architectures» Distributed Parallel processing Vs Clustered Parallel processing

» Should Von-Neumann and Harvard Architectures be retired

» Summary


Introduction

» FPGAs………a 20 year History!» From “Glue Logic” beginnings to complete

“Systems on a Chip” today. » Mathematical functions in FPGAs since 1993 » FPGAs pervasive in virtually all electronics

equipment » Many different perspectives on FPGA capability» This leads to confusion in the market place and

mixed messages


FPGA Perceptions

ASICMicro-

processor

Easy to ProgramFlexible

Difficult to programFixed Function

Low Performance High Performance

DSP FPGA

When an FPGA is viewed as an ASIC……………..…the observer historically saw……

Low performance

Wasted TransistorsHigher Power Consumption

Good for prototyping

Never Use in Production

High Cost


FPGA Perceptions

ASICMicro-

processor




DSP FPGA

These were the original views of FPGAs and they have STUCK in many peoples minds

Low performance

Wasted TransistorsHigher Power Consumption

Good for prototyping

Never Use in Production

High Cost


FPGA Perceptions

ASICMicro-

processor




DSP FPGA

When an FPGA is viewed as a DSP……………..…today the observer sees ……

Co-ProcessorHigh Performance

I/O InterfaceNiche


FPGA Perceptions

ASICMicro-

processor




DSP FPGA

When an FPGA is viewed as a processor……………..…Nallatech sees ……

Main ProcessorHigh Performance

Floating Point

Lower Power Consumption

Increased Performance Density

Immature Tools


Have It Your Way!

ASICMicro-

processor




FPGA


HPEC Vs HPC

» Earlier FPGA adoption within HPEC Community » FPGA based HPEC is a volume based commercial

reality today » High Performance Embedded Computing (HPEC)

» Users have an appreciation of underlying hardware technology» Low level of programming abstraction» Applications are often severely SWAP restricted

» High Performance Computing (HPC)» Users are far more software centric » Programming is achieved using high level software languages» Exclusive use of Floating Point arithmetic» Focus on ease of implementation of complex algorithms


1993 – Simple Maths Functions within FPGA

» Simple Mathematical Functions» 2 bit arithmetic function per Logic Slice» Effective for 1D Pipelined data paths» Very Basic Functions - highly repetitive and Data Intensive» Schematic Hardware designed

» XC4006 Series FPGAs

» 256 Logic Slices» 128 max user I/O


MicroprocessorOnly

Microprocessorbased

FPGAcoprocessor

1990

Pre Nallatech

1993

Professional Consulting Services

Nallatech’s Adoption of FPGAs for HPEC


Nallatech’s Early FPGA HPEC Computing Experience - 1993

» Real time, ultra low latency, Imaging Simulator» Embedded Distributed Processing System» Floating point, matrix transformations, convolution, sensor interfacing etc

Gain + Offset Control Convolution

Image Composition

Sensor Interface

Target Image Generation

Background Image Processing

T9FPGA I860

x6 x4x2I860I860I860T9

FPGAFPGAFPGAFPGAFPGA

FPGAx2 FPGA

T9

FPGA

x3

x1

T9T9

T9

FPGA

C80

x2

x3

x4

T9

FPGAFPGA

C80C80C80

T9x3 T9T9


1998 – Virtex FPGA Family

» Revolutionary Xilinx Virtex FPGA family Introduced» 32 x 4Kbits Block RAMs + other mathematical features

introduced» allowed 2D mathematical algorithms to be implemented» Excellent for Image Processing Algorithms

» Significant DSP capability

» Virtex

» 12,000 Logic Slices» 804 max user I/O» 32 4Kbit Block RAMs


MicroprocessorOnly

Microprocessorbased

FPGAcoprocessor

1990

Pre Nallatech

1993



Microprocessor coprocessor

FPGAbased

1998DIME Product

Family FPGA Centric

Computing Architecture


2001 – Virtex-II FPGA Family

» Virtex-II FPGA introduced followed by Virtex-II Pro in 2003» 444 18x18 Multipliers & 18kbit block RAMs introduced» Gbit Serial I/O Communications & Power PC Processors Introduced» Complex Floating Point Algorithm Implementation now possible

» Virtex-II / Pro» 44,000 Logic Slices» 444 18Kbits BRAMs» 444 18x18

Multipliers» 2 PowerPC

Processors» 20 Gbit I/O» 1164 Max User I/O


What This Means for HPC

MicroprocessorItanium 2

FPGAVirtex 2VP100

Technology 0.13 Micron 0.13 Micron

Clock Speed 1.6GHz 180MHz

Internal Memory Bandwidth

102 GBytes per Sec 7.5 TBytes per Sec

# Processing Units

5 FPU(2MACs + 1FPU) + 6 MMU + 6 Integer Units

212 FPU or 300+ Integer Units

or ……….

Power Consumption

130 WATTS 15 WATTS

Peak Performance 8 GFLOPs 38 GFLOPS

Sustained Performance

~2 GFLOPs ~19 GFLOPS

I/O / External Memory Bandwidth

6.4 GBytes/sec 67 GBytes/sec


MicroprocessorOnly

Microprocessorbased

FPGAcoprocessor


FPGAbased

1990

Pre Nallatech

1993


1998

DIME Product Family

FPGA Centric Computing Architecture


FPGA

Microprocessor embedded

2001

DIME-II Product Family



HardwarePlatform

System Software

SystemsCommunications

FPGA ComputingThe Whole Solution

» Inter-FPGA Communication

»Abstracts Hardware Architecture

» System Management and control

» APIs

» COTS Hardware

» Modular

» Multiple-FPGA Systems


FPGA Communications and Tool Support

From PCI

Physical Link

FPGA

PCI Host

FPGA

N3

N4

R

N5

R N6N2

N1

RE

BB

VHDL

Memory

MATLAB

Processors

Block Flows

C FlowsOpen 3rd Party Component

Support

N

N N

N N

N

Viva

- C


Accelerated Hardware Implementation

From PCI

Physical Link

FPGA

PCI Host

FPGA

N3

N4

R

N5

R N6N2

N1

RE

BB

VHDL

Memory

MATLAB

Processors

Block Flows

C Flows3rd Party Component

Support

N

N N

N N

N

Viva

- C


Commercial Realities for HPC

» Not a Panacea – As with all parallelisation» Translation of Legacy Code

Legacy Code

C Program

FPGA Translated

Execution Time

0%

100%

On Processor

Bandwidth & Latency

Considerations

uP ExecutedFPGA

Executed

uP / FPGA Partition

Execution Time



» Maturity of High Level Languages» Good progress has been made in FPGA compilers» Often Trade off between performance and ease of use

» Parallelising of code» Fine Grain parallelism is critical» Still not automatic» Taking implied parallelism from serial code is NOT good

enough» HPC Software engineers are well qualified to deal with this

» Development & Debug Time» Comparable to programming in assembler vs C» Biggest Hurdle is the Synthesis times –hours to days!» Where tools can make a significant impact- If they have no

bugs!



» No Real Industry Standardisation » Requires expertise to “brew your own solutions”» Difficulty for Beginners

» Bang for Buck – for Floating Point Implementations» NRE today WILL be more expensive - ~5-10 times» Can approach 100 times performance for 1/2th the Cost » Significantly reduced SWAP, >200 times,

» Result in a significant Cost of Ownership savings



MicroprocessorOnly

Microprocessorbased

FPGAcoprocessor


FPGAbased

1990

Pre Nallatech

1993


1998

DIME Product Family


FPGA

Microprocessor embedded

2001

DIME-II Product Family


FPGA

2003

FPGA Based HPC Solutions


Algorithm Acceleration

Seismic Processing - Kirchhoff algorithm

- Single Precision Floating Point - 64 times faster than a 2GHz Pentium 4-200 times less power consumption

Smith-Waterman- Dynamic Programming Algorithm used in Biological Sequencing- 155 times faster than SunFIRE 280R processing unit


Algorithm Acceleration

Real Time Video Processing - Single Precision

Floating Point calculations-36 GFlops + 40 GOPs sustained Performance on a single PCI card- >200 times Power reduction over Xeon

Gravity Simulation

- N-Body computation- Single Precision Floating Point - 20GFlops/sec sustained performance-100 times faster than 2.4GHz Pentium 4 CPU


A Young Solar System


Colliding Galaxies


FPGA CoprocessorVs

FPGA Main Processor


FPGA’s as Coprocessors

» Accelerator / Offload Engine for Microprocessor based solutions

» Advantages» Easy to conceptualise» Pragmatic Approach» Possible large performance improvement for least effort» Port small functions of Compute intensive Legacy Code» Rest of code remains on Existing Host. » Benefit from Existing Host interfaces

» Disadvantages » Only Applicable to certain functions» Need to consider bandwidth / latency requirements


FPGA’s as Main Processor

» The FPGA takes on the complete Compute Function

» Advantages» Build the Computing Architecture around Algorithmic

Problem» Can provide another order of magnitude increase in

performance» Can go back to First principles» Don’t have to port Optimised processor code to FPGA

» Disadvantages» Rarely Start with a clean sheet of paper» Tool Maturity» Only practical for relatively straight forward algorithms


Main Vs Co - Recommendations

» Co-processor approach is most applicable Today : -» For HPC» Whenever a Man Machine interface is required» Whenever low performance industry standard interfaces are

required. » If you need to work with legacy code» Quick wins

» A Main Processor Approach is recommended Today : -» For Stand alone Embedded applications (HPEC)» When starting with a clean sheet of paper» If ultimate performance is a pre-requisite » The best power/performance ratio is required from your system» Relaxed development times


Optimizing Spatial & Temporal Demands

of Computing Problems


Spatial & Temporal Definitions

» There are several perspective on the meaning of spatial and temporal 1. Cluster of Microprocessors

» Temporal = Function runs within one processor» Spatial = Function spread across many microprocessor nodes

2. Traditional Embedded Computing Hardware» Temporal = using Microprocessor» Spatial = Implementing dedicated ASIC accelerators

3. FPGAs» Temporal = using same logic resources for multiple Functions» Spatial = Paralleling and pipelining a function across the FPGA

Fabric

» Ultimate Aim is to ensure that no processor goes idle» And you utilise all the available resources


Complexity and Speed

» Any Application will be constructed from several functions» Each Function will have varying degrees of complexity» Each Function will also have varying demands on its

execution time

ComplexSimple

Can Execute Slowly

Must Execute Quickly

ComputeIntensive

MediumCompute

LowCompute

MediumCompute

Fully spatial implementati

on

Balanced spatial and temporal

implementation

Fully Temporal

Implementation

TraditionalASIC

Traditional Vector

TraditionalMicroproces

sorFPGA

All parallel / pipelined

FPGAPartially Parallel

Partial reuse

FPGASoft

Microprocessor


Homogenous ComputingVersus

Heterogeneous ComputingVersus

Polymorphic Computing


Direction of Computing

Cell 1

PPC

Cell 3

Cell 5

Cell 7

Cell 2

Cell 4

Cell 6

Cell 8

Cell Processor – IBM, Sony, Toshiba

Heterogeneous Computing on Silicon

Global Shared Memory

uPuP

GP

FPGAFPGAASSP

SGI – Heterogeneous Architecture

Intel – 16 processor per dieHomogeneous on Silicon




Application Data Types

SymbolicVector/StreamingBit Level

SW

EP

T

Effi

cie

ncy

Siz

e, W

eig

ht,

Energ

y, Perf

orm

ance

, Tim

e

FPGA Processor

Polymorphic


Cell

Pro

cess

or

RP –

e.g

. C

lears

peed

Elix

ent,

Pic

och

ip

Support

for

FPU

?

Support

for

DSP


Application Data Types

SymbolicVector/StreamingBit Level

SW

EP

T

Effi

cie

ncy

Siz

e, W

eig

ht,

Energ

y, Perf

orm

ance

, Tim

e

FPGA Processor


What is Polymorphic processor?

MicroprocessorDSP UnitFloating Point Operator

Integer OperatorLogical Operator


A Polymorphic FPGA

» FPGA is the closest concept to Polymorphic computing» It can morph into the different operators» However it cannot perform them all with equal efficiency

» Is a polymorphic course grained FPGA Possible

= Polymorphic processing elementCan Morph into : -•Microprocessor•Integer Operator•DP/SP Floating Point Operator•Logical Operator•Text Operator

= Traditional FPGA Fabric


Coarse Grain ArchitecturesVersus

Fine Grain Architectures


Coarse vs Fine Grain

» Cluster = ultimate in coarse grained parallelism» ASIC = Ultimate in fine grain parallelism» FPGA = Programmable Fine grain parallelism» The Finer the grain, the more you can make the

architecture exactly fit the problem.» However Fine Grain Programmable FPGA are a

sub-optimal solution as they suffer from» An inefficient transistor utilisation on coarser grain

operations» A slower clock frequency that could be improved with

coarser granularity


Distributed Parallel Processing (DPP)

VsCluster Parallel Processing

(CPP)


DPP & CPP Definitions

» Processing Power is distributed to where it is needed

» Direct Communications built as needed

» Computing Architecture designed to fit the Application


Image Composition

Sensor Interface



T9FPGA I860

x6 x4x2I860I860I860T9


FPGAx2 FPGA

T9

FPGA

x3

x1

T9T9

T9

FPGA

C80

x2

x3

x4

T9

FPGAFPGA

C80C80C80

T9x3 T9T9


Image Composition

Sensor Interface




Image Composition

Sensor Interface



T9FPGA I860

x6 x4x2I860I860I860T9


FPGAx2 FPGA

T9

FPGA

x3

x1

T9T9

T9

FPGA

C80

x2

x3

x4

T9

FPGAFPGA

C80C80C80

T9x3 T9T9

T9FPGA I860

x6 x4x2I860I860I860T9

FPGAFPGAFPGAFPGAFPGAT9

FPGA I860

x6 x4x2I860I860I860T9


FPGAx2 FPGAFPGAx2 FPGA

T9

FPGA

x3

x1

T9T9T9

FPGA

x3

x1

T9T9

T9

FPGA

C80

x2

x3

x4

T9

FPGAFPGA

C80C80C80

T9

FPGA

C80

x2

x3

x4

T9

FPGAFPGA

C80C80C80

T9x3 T9T9

T9x3 T9T9

CommunicationsInfrastructure

= Server node

Distributed Parallel ProcessingCluster Parallel Processing

» Regular processor Architecture » Regular communications

Infrastructure» Application must be designed

to fit the computer architecture


Application Implementation

» Example application consisting of 8 algorithms

» Need to map onto hardware for real-time implementation

» Algorithms each have different characteristics

Algorithm A

Algorithm B

Algorithm C

Algorithm D

Algorithm E

Algorithm F

Algorithm G

Algorithm H


CommunicationsInfrastructure

Fitting Application to Cluster

Algorithm A

Algorithm B

Algorithm C

Algorithm D

Algorithm E

Algorithm F

Algorithm G

Algorithm H

Application Cluster


Fitting Application to Distributed FPGA Computer

» VME Blade form-factor» Five high-density platform

FPGAs» High-speed external

analog interfaces» High-speed synchronous

SRAM memory» Gigabit Ethernet interface


Application Implementation

» Example application consisting of 8 algorithms

» Need to map onto hardware for real-time implementation

» Algorithms each have different characteristics

Algorithm A

Algorithm B

Algorithm C

Algorithm D

Algorithm E

Algorithm F

Algorithm G

Algorithm H


Same Application Implemented on FPGAs

FPGA

FPGA

VME FPGA

GBit E

thern

et

FPGA

FPGA

C(PicoBlaze uP)

Algorithm B

VHDL

Algorithm AAlgorithm D

C(MicroBlaze

uP)Algorithm CAlgorithm F

Verilog

Algorithm G

MATLAB or Simulink

Algorithm HAlgorithm E


Communications network to connect algorithms

FPGA

FPGA

VME FPGA

GBit E

thern

et

FPGA

FPGA

B

N

N

B B

E R

R

N

N

B

B

B

N

N

B

N

N

N

R

N

R

B

R

N

R


So, should Von-Neuman and Harvard Architectures be

retired?


Should they Retire?

» Von Neumann and Harvard provide highly efficient use of silicon real estate whilst still being capable of executing any computational function.

» Therefore perhaps they should still live on» However this will be less and less in a hard chip

implementation» The Intelligent compiler will instantiate a Von-

neumann or Harvard like architecture when they are the most efficient way to execute an algorithm

Von-Neumann and Harvard will live on as part of the intelligence within tomorrow’s Compilers.


Summary

» FPGAs for computing is not new» 12 Years accelerating maths functions» Floating Point & Tools make FPGAs viable

for HPC Community» No coherent Industry Standardisation » Code development WILL take longer» Significant potential savings

» Price/Performance» SWAP» Cost of Ownership


And Finally……………

SGI & Nallatech have formally agreed on a Strategic

Collaborative ArrangementThis brings together 12 years of expertise in delivering Real FPGA computing solutions from Nallatech with the Global Shared Memory MPP

computing from SGI.Customers now have a path to scale from a

commodity cluster with FPGAs all the way up to a massive HPC system with thousands of

Processors and & thousands of FPGAs


Thank You for your attention

www.nallatech.com

Copyright © 2005 Nallatech Limited. All rights reserved. Nallatech, the Nallatech logo, the triangles device and “The High Performance FPGA Solutions Company" are trademarks of Nallatech Limited. All other trademarks acknowledged.

commercial in confidence. copyright ©2005, nallatech.1 is it time for von neumann and harvard to...

Documents

fpga perceptionsfpgawhen

fpga capabilitythis

fpga perceptionsthese

hpec community fpga

nallatechs adoption

fpgas pervasive

d mathematical algorithms

logic sliceeffective