Model-Driven Physical-Design for Future Nanoscale Architectures

Model-Driven Physical-Design for Future Nanoscale Architectures

Ciprian TEODOROV

Lab-STICC MOCS

November, 28th 2011

Generic physical-design framework

based on a common vocabulary

is the key to taming

nanoscale architectures.

Thesis statement

Context

• Smaller & denser circuits which consume much less power

Society Needs

• Current technology (CMOS) reaches its limits

However

• Different architectural propositions based on emergent technologies

State of the art

Model-Driven Physical-Design for Future Nanoscale Architectures

3

Examples

Quantum

• quantum-dot cellular automata

Molecular

• Tour’s Nanocell

Crossbar

• NanoPLA, CMOL, NASIC, FPNI

Model-Driven Physical-Design for Future Nanoscale Architectures

4

What is a Crossbar ?

CrosspointDifferent devices: • Diode• FET• Etc.

Model-Driven Physical-Design for Future Nanoscale Architectures

5

Example: NASICutilization and performance.

5. NASIC Architectures Large scale computing systems may be designed using the NASIC fabric and the associated

framework of building blocks, xnwFET based circuits, and logic styles discussed in the

preceding sections. This section discusses two key architectures for the NASIC fabric: the WISP-

0 general purpose processor and a massively parallel architectural framework with

programmable templates for image processing.

5.1 WIre Streaming Processor (WISP-0)

WISP-0 is a stream processor that implements a 5-stage

microprocessor pipeline architecture including fetch,

decode, register file, execute and write back stages.

WISP-0 consists of five nanotiles: Program Counter (PC),

ROM, Decoder (DEC), Register File (RF) and Arithmetic

Logic Unit (ALU). Figure 17 shows its layout. A nanotile

is shown as a box surrounded by dashed lines in the

figure. In WISP designs, in order to preserve the density

advantages of nanodevices, data is streamed through the

fabric with minimal control/feedback paths. All hazards

are exposed to the compiler. It uses dynamic circuits and

pipelining on the wires to eliminate the need for explicit

flip-flops and therefore improve the density considerably.

WISP-0 supports a simple instruction set including nop,

mov, movi, add and multiply functions. It uses a 7-bit

instruction format with 3-bit instruction and 2-bit source

and destination addresses. The WISP-0 is used as a design

prototype for evaluating key metrics such as area and

performance as well as the impact of various fault-

tolerance techniques on chip yield and process variation

mitigation. Additional enhancements to this design are

ongoing in the NASIC group.

5.1.1 WISP-0 Program Counter

The WISP-0 program counter is implemented as a four bit accumulator. Its output is a four bit

address that acts as an input to the ROM. The address is incremented each cycle and fed back

using a nano-latch. Figure 18 shows the implementation of the Program Counter with a NAND-

NAND scheme. Diagonal transistors on the upper two NAND planes implement the nano-latch

to delay the output by one cycle and allow the signals to ‘turn the corner’.

Figure 18. WISP-0 Program Counter

Figure 17. WISP-0 Processor Floorplan

Model-Driven Physical-Design for Future Nanoscale Architectures

6

Common Features of Crossbar Fabrics

Nanowire based

Regularity of assembly leads to:

• a crossbar-like structure

• PLA and/or FPGA-like fabric architecture

CMOS superstructure, thus a nano-CMOS interface

Large number of defects

Logic implementation

Model-Driven Physical-Design for Future Nanoscale Architectures

7

Differences Between Crossbar Fabrics

Architectural differences

• Nano-role, CMOS-role

• Fabrication strategy

Physical parameters used for evaluation

• CMOS/NW pitch / device characteristics

Evaluation strategies

• Yield simulation

• Place & Route on predefined array

Hypotheses during evaluation

• Defect/Faults models

Model-Driven Physical-Design for Future Nanoscale Architectures

8

Research Questions

How to maximize the reuse of design-tools?

Is it possible to create a generic design toolkit?

How to separate the algorithmic and architectural concerns?

How to add a tools axis to design-space exploration problem?

How to integrate multi-level fault tolerance?

Model-Driven Physical-Design for Future Nanoscale Architectures

9

Contributions

Model-driven physical-design @ nanoscale

Common vocabulary

Tools as model transformation

Reified design flow

DSE bootstrap methodology

R2D NASIC: Nanoscale

architecturetemplate

Max-rate pipeline routing

Model-Driven Physical-Design for Future Nanoscale Architectures

10

Outline

MoNaDe Toolkit

•Overview

•Structural Domain Modeling

•Tool Modeling

•Design flow Modeling

R2D NASIC

•Overview

•Analytic evaluation

•Characteristics

R2D NASIC Evaluation Results

•Surface / Performance

•Max-rate pipeline evaluation

•Room for improvements

Conclusion & Perspectives

Model-Driven Physical-Design for Future Nanoscale Architectures

11

Outline

MoNaDe Toolkit

•Overview

•Structural Domain Modeling

•Tool Modeling

•Design flow Modeling

R2D NASIC

•Overview

•Analytic evaluation

•Characteristics

R2D NASIC Evaluation Results

•Surface / Performance

•Max-rate pipeline evaluation

•Room for improvements

Conclusion & Perspectives

Model-Driven Physical-Design for Future Nanoscale Architectures

12

System

RTL

Circuit

Fabric

Device

Model-Driven Physical-Design for Future Nanoscale Architectures

13

Typical Toolkit

Model-Driven Physical-Design for Future Nanoscale Architectures

14

Defect/Faults Application

Architecture ToolsPacking

Placement

Routing

UI

P&R OutputMetrics

MoNaDe

Structural Domain Modeling

Common abstract model

• Hierarchical annotated port graph

• Specialized to model applications and architectures

• Provides a common vocabulary (Entities and API)

• Enables the creation of generic utilities

Model-Driven Physical-Design for Future Nanoscale Architectures

15

Hierarchical Annotated Port Graph

Model-Driven Physical-Design for Future Nanoscale Architectures

16

Structural Modeling:NASIC case

Model-Driven Physical-Design for Future Nanoscale Architectures

17

Model-Driven Physical-Design for Future Nanoscale Architectures

18

Structural Modeling:NanoPLA

Model-Driven Physical-Design for Future Nanoscale Architectures

19

Tools as Transformations

Decouple the algorithms from the domain models

The tools are implemented as composite model-to-model transformations

The tools refine the domain models

Integration of external tools and algorithms

Model-Driven Physical-Design for Future Nanoscale Architectures

20

Transformation Meta-Model

Model-Driven Physical-Design for Future Nanoscale Architectures

21

Model-Driven Physical-Design for Future Nanoscale Architectures

22

Architectural Model

RRGraph Extraction

Application Model

Nets ExtractionTGraph Extraction

Placed App Model

Nets

Pathfinder

RRGraph

Post-routing FPGA Model Refinement

TGraph

add AT add RTupdate costs

Routes

add routes

Tool-Flow Modeling

Reified tool-flow as a composite transformation

The physical-design tool-flow is a DAG of tools

Capacity to create tool-flow derivations

Enables incremental tool-flow creation

Enables Architecture/Tools exploration

Model-Driven Physical-Design for Future Nanoscale Architectures

23

Tool-Flow Hierarchy Example

Model-Driven Physical-Design for Future Nanoscale Architectures

24

SimpleDirectedAlongRoutePathEqFlow

NCellBasedSynthesis

VPRBasedFlow

LagrangianTimingDrivenFlow

DirectPlaceRouteDirectedConnections

VPRDirectPlaceAndRoute

DirectPlaceRoute

LagrangianPathEqualisationFlow

NAbstractFlow

CompleteArrayDirectPR

SimpleDirectedFSSwitchPathEqFlow

NFlowBasedSynthesis

MadeoBasedFlow

SimpleFSwitchPathEqualisationFlow

BASELINE

MAX-RATE

Outline

MoNaDe Toolkit

•Overview

•Structural Domain Modeling

•Tool Modeling

•Design flow Modeling

R2D NASIC

•Overview

•Analytic evaluation

•Characteristics

R2D NASIC Evaluation Results

•Surface / Performance

•Max-rate pipeline evaluation

•Room for improvements

Conclusion & Perspectives

Model-Driven Physical-Design for Future Nanoscale Architectures

25

NASIC 2D Routing Problem

Model-Driven Physical-Design for Future Nanoscale Architectures

26

Model-Driven Physical-Design for Future Nanoscale Architectures

27

R2D NASIC – Routing & Timing

Model-Driven Physical-Design for Future Nanoscale Architectures

28

Results – NW Length

Model-Driven Physical-Design for Future Nanoscale Architectures

29

Performance vs Input width

0,00E+00

2,00E+08

4,00E+08

6,00E+08

8,00E+08

1,00E+09

1,20E+09

1,40E+09

1,60E+09

1,80E+09

1 3 5 7 9111315171921232527293133353739414345474951535557596163656769

Frequency(Hz)

#ofinputs

Model-Driven Physical-Design for Future Nanoscale Architectures

30

R2D NASIC – Main Characteristics

Compatibility with NASIC technology and fabrication

Adaptability to technological and applicative constraints

Compatibility with NASIC fault-tolerance techniques

Custom placement and routing due to structural regularity

Max-rate pipeline designs due to pipelined routing architecture

Simplified delay estimation

Model-Driven Physical-Design for Future Nanoscale Architectures

31

Outline

MoNaDe Toolkit

•Overview

•Structural Domain Modeling

•Tool Modeling

•Design flow Modeling

R2D NASIC

•Overview

•Analytic evaluation

•Characteristics

R2D NASIC Evaluation Results

•Surface / Performance

•Max-rate pipeline evaluation

•Room for improvements

Conclusion & Perspectives

Model-Driven Physical-Design for Future Nanoscale Architectures

32

R2D NASIC Toolflow

Model-Driven Physical-Design for Future Nanoscale Architectures

33

Application

Architecture Tools PLAMap

Placement

RoutingUI

P&R OutputMetrics

MoNaDe

PLA Explore

SIS

UI

Normalized Density Gain over 45nm standard cell

17X

1.32X

48X

2X

24X

12X9X

1

10

100

alu4 apex2 apex4 des ex5p misex3 seq

Model-Driven Physical-Design for Future Nanoscale Architectures

34

BASELINE

Speed

43MHz67MHz

40MHz

9MHz

167MHz

29MHz 27MHz

1

10

100

1000

alu4 apex2 apex4 des ex5p misex3 seq

Frequency

Operating frequency of the slowest logic stage / throughput

Results assume 1GHz for the slowest logic stage

Too slow

Model-Driven Physical-Design for Future Nanoscale Architectures

35

BASELINE

Max-Rate Pipeline System

Add REs

Model-Driven Physical-Design for Future Nanoscale Architectures

36

Model-Driven Physical-Design for Future Nanoscale Architectures

37

R2D NASIC Toolflow

Model-Driven Physical-Design for Future Nanoscale Architectures

38

Application

Architecture Tools PLAMap

Placement

RoutingUI

P&R OutputMetrics

MoNaDe

PLA Explore

SIS

UI

Max-Rate Router

Net Performance Improvement

Model-Driven Physical-Design for Future Nanoscale Architectures

39

MAX-RATE

Normalized Density Gain over 45nm standard cell

17X

1.32X

48X

2X

24X

12X9X

3X

0.03X

12X

0.06X

13X

1.24X

0.46X

0.01

0.1

1

10

alu4 apex2 apex4 des ex5p misex3 seq

Baseline Max-Rate

Model-Driven Physical-Design for Future Nanoscale Architectures

40

MAX-RATE

Performance*Area Gain

66X

0.32X

274X

5X

61X40X

14X

0.1

1

10

100

1000

alu4 apex2 apex4 des ex5p misex3 seq

Model-Driven Physical-Design for Future Nanoscale Architectures

41

MAX-RATE

Room for Improvement: STDEV of Routing Block Usage

Model-Driven Physical-Design for Future Nanoscale Architectures

42

MAX-RATE

Outline

MoNaDe Toolkit

•Overview

•Structural Domain Modeling

•Tool Modeling

•Design flow Modeling

R2D NASIC

•Overview

•Analytic evaluation

•Characteristics

R2D NASIC Evaluation Results

•Surface / Performance

•Max-rate pipeline evaluation

•Room for improvements

Conclusion & Perspectives

Model-Driven Physical-Design for Future Nanoscale Architectures

43

Conclusions

Generic physical-design toolkit for nanoscale crossbar fabrics

• Model-driven approach: structure, algorithmics, and flow reified.

• Two main abstraction levels considered

• Quantitative incremental DSE, bootstrapped with standard tools + new exploration axis

Nanoscale architecture template

• Enables arbitrary routing for NASIC

• Shows the impact of pipelined (dynamic logic) routing

Model-Driven Physical-Design for Future Nanoscale Architectures

44

Future Research

MoNaDe toolkit

• Formalize the models and the API

• Hardware accelerated physical-design using external transformations

• Create an optimizing tool-flow execution engine

• Open infrastructure for architectural exploration in the context of new technologies

Nanoscale architectures @ techno level

• Fault tolerance for dynamic routing

• Parameter variability impact on multi-tile design

• Clock distribution in highly constrained 2D topologies

Nanoscale architecture @ design level

• Creation and/or improvement of pipeline aware tools (placement, routing, etc)

• Study the extent to which dynamic logic evaluation impacts the physical design for other architectures besides R2D NASIC

Model-Driven Physical-Design for Future Nanoscale Architectures

45

Model-Driven Physical-Design for Future Nanoscale Architectures

46

Configuration Management

Configuration types:

• Structural

• Fine-grain functional

• Coarse-grain functional

Configuration state-machine reified in the model

Different configuration policies

• One-time configuration

• Reconfiguration

Model-Driven Physical-Design for Future Nanoscale Architectures

47

Fault-Modeling and Injection

Domain concepts specialized to faulty entities

• FET -> StuckAt0FET and StuckAt1FET

Two ways to inject faults:

• Object swapping: injects faulty devices by replacing model instances

• Fault-configuration: uses a probabilistic configuration controller

Model-Driven Physical-Design for Future Nanoscale Architectures

48

Fault-Modeling and Injection - Details

Model-Driven Physical-Design for Future Nanoscale Architectures

49

NFET StuckAt1

StuckAt0

NULL

30%

NFET StuckAt1

StuckAt0

NULL30%

10%

60%