+ All Categories
Home > Documents > 1 3/22/02 Benchmark Update u Carnegie Cell Library: “Free to all who Enter” s Need to build...

1 3/22/02 Benchmark Update u Carnegie Cell Library: “Free to all who Enter” s Need to build...

Date post: 21-Dec-2015
Category:
View: 215 times
Download: 0 times
Share this document with a friend
Popular Tags:
17
3/22/ 02 1 Benchmark Update Benchmark Update Carnegie Cell Library: “Free to all who Carnegie Cell Library: “Free to all who Enter” Enter” Need to build scaling model of standard cell Need to build scaling model of standard cell library library Based on our open 0.35 micron library (real Based on our open 0.35 micron library (real extracted data) extracted data) This semester: basic standard cells This semester: basic standard cells This summer: memories This summer: memories Timing Models Timing Models Actual timing probably is not as important as Actual timing probably is not as important as variations variations Simple 1-order models of speed and Simple 1-order models of speed and variations variations
Transcript

3/22/02 1

Benchmark UpdateBenchmark Update

Carnegie Cell Library: “Free to all who Enter”Carnegie Cell Library: “Free to all who Enter” Need to build scaling model of standard cell libraryNeed to build scaling model of standard cell library

Based on our open 0.35 micron library (real extracted data)Based on our open 0.35 micron library (real extracted data)

This semester: basic standard cellsThis semester: basic standard cells

This summer: memoriesThis summer: memories

Timing ModelsTiming Models Actual timing probably is not as important as variationsActual timing probably is not as important as variations

Simple 1-order models of speed and Simple 1-order models of speed and variationsvariations

3/22/02 2

Circuit BenchmarksCircuit Benchmarks

New vertical benchmarks:New vertical benchmarks:All designed to comply with common network interfaceAll designed to comply with common network interface Interoperability, portability for IP blocksInteroperability, portability for IP blocksWhy is this interesting/useful for benchmarks?Why is this interesting/useful for benchmarks?

IOs and other system-level issues make it hard to compare benchmarksIOs and other system-level issues make it hard to compare benchmarks

Also good for educationAlso good for education

NetworkTile

ProcessingElement

NetworkTile

ProcessingElement

NetworkTile

ProcessingElement

NetworkTile

ProcessingElement

NetworkTile

ProcessingElement

NetworkTile

ProcessingElement

3/22/02 3

New Circuit BencmarksNew Circuit Bencmarks

The Network Tile: for streaming applicationsThe Network Tile: for streaming applications

The Processing Elements:The Processing Elements: Morphable Floating Point Multiplier: Morphable Floating Point Multiplier:

FP mult and vector add, integer multiply and integer MAC and shiftFP mult and vector add, integer multiply and integer MAC and shift

Morphable Floating Point Adder:Morphable Floating Point Adder: FP add and integer add and shiftFP add and integer add and shift

Programmable Integer ALUProgrammable Integer ALU Programmable FIR filterProgrammable FIR filter SIMD Adder (with funky completion logic)SIMD Adder (with funky completion logic)

All about 20-100k gates eachAll about 20-100k gates each Can be combined into systems of arbitrary sizeCan be combined into systems of arbitrary size Network limits effective Rent’s ExponentNetwork limits effective Rent’s Exponent *Actually Network connectivity would determine Rent Exponent*Actually Network connectivity would determine Rent Exponent

Currently planning 2-D network, creating Rent Exponent of 0.5Currently planning 2-D network, creating Rent Exponent of 0.5

3/22/02 4

Dynamic Network TilesDynamic Network Tiles

PE

TILE

PE

TILE

PE

TILE

PE: the component at this nodein the network

TILE: the network component

3/22/02 5

Dynamic Network SystemDynamic Network System

PE

TILE

PE

TILE

PE

TILE

PE

TILE

PE

TILE

PE

TILE

PE

TILE

PE

TILE

PE

TILE

PE

TILE

PE

TILE

PE

TILE

Reference Clock distribution through network

Each tile generates own clockInterface decoupled via FIFOs

New Placement Problem:space utilization vs. distance

3/22/02 6

Target Architecture:Target Architecture:

Pipelined ArraysPipelined Arrays

Limited FeedbackLimited Feedback

Long/short wires predictableLong/short wires predictable

Clock SkewClock Skew

Important Application DomainImportant Application Domain

3/22/02 7

Classic Wire Length ModelsClassic Wire Length Models

D

A

C

B

D

A

C

B

3/22/02 8

Wire Path LengthWire Path Length

Every block is a pipeline stageEvery block is a pipeline stage

Impossible to determine every wire length from floorplan Impossible to determine every wire length from floorplan

blocksblocks

Wire Path Length (WPL) measures the distance between Wire Path Length (WPL) measures the distance between

consecutive registersconsecutive registers

3/22/02 9

Wire Path LengthWire Path Length

D

A

C

B

D

A

C

B

3/22/02 10

Results KeyResults Key

ClassicClassic Different random starting position every timeDifferent random starting position every time

Classic Move Set - SwapClassic Move Set - Swap

Classic + LSPClassic + LSP Same legal starting position every timeSame legal starting position every time

Classic Move Set - SwapClassic Move Set - Swap

NewNew Same legal starting position every timeSame legal starting position every time

New Move Set - Insert/DeleteNew Move Set - Insert/Delete

3/22/02 11

IDEA 60 Block DesignIDEA 60 Block Design

0.00%

50.00%

100.00%

150.00%

200.00%

250.00%

0.00% 2.00% 4.00% 6.00% 8.00% 10.00%Area Greater than Minimum

Wir

e P

ath

Len

gth

Lo

ng

er t

han

M

inim

um

Classic

Classic + LSP

New

3/22/02 12

1-D DCT1-D DCT

12 Pipeline Stages12 Pipeline Stages

Synthesis Speed - 2.25 ns.Synthesis Speed - 2.25 ns.

Synthesis Area - 668,323 Synthesis Area - 668,323 mm22

3/22/02 13

UnfloorplannedUnfloorplanned

3/22/02 14

ClassicClassic

3/22/02 15

Classic + LSPClassic + LSP

3/22/02 16

NewNew

3/22/02 17

ResultsResults

DeadSpace

WPL DRCs SpeedAvg.

Cong.Max

Cong.

NoFloor X X 0 2.50 0.35 0.98

Classic 0.97% 6.15 850 2.40 0.60 1.32

Classic+ LSP 1.63% 5.88 138 2.44 0.57 1.31

New 0.09% 5.76 12 2.37 0.54 1.19


Recommended