isca needle a_0610

64
Moving the Needle Computer Architecture Research in Academe and Industry ll Dally ief Scientist & Sr. VP of Research, NVIDIA ll Professor of Engineering, Stanford University

Upload: parallellabs

Post on 15-Jan-2015

744 views

Category:

Technology


1 download

DESCRIPTION

Moving the Needle: Computer Architecture Research in Academe and IndustryBy Bill Dally

TRANSCRIPT

Page 1: Isca needle a_0610

Moving the NeedleComputer Architecture Research

in Academe and IndustryBill DallyChief Scientist & Sr. VP of Research, NVIDIABell Professor of Engineering, Stanford University

Page 2: Isca needle a_0610

Outline

The Research FunnelMost ideas fail

Those that succeed take 5-10 years

The Research Formula

Constraints

The Academic Advantage

The Industrial Advantage

Startups

Best practices

Page 3: Isca needle a_0610

Goal – Positive Impact on a Product

Page 4: Isca needle a_0610

The Research Funnel

Applications

Technology

ConceptDev

Model Eval Dev

insight

Page 5: Isca needle a_0610

Most ideas fail

The ideas that succeed take a long time

ConceptDev

Model Eval Dev

Page 6: Isca needle a_0610

Most ideas fail

The ideas that succeed take a long time

ConceptDev

Model Eval Dev

Page 7: Isca needle a_0610

Most ideas fail

So terminate the bad ones quickly

Page 8: Isca needle a_0610

Most ideas fail

So terminate the bad ones quickly

Be a terminator, not an advocate

Page 9: Isca needle a_0610

Dally, “Micro-Optimization of Floating-Point Operations, ASPLOS, 1989, pp 283-289

Page 10: Isca needle a_0610
Page 11: Isca needle a_0610

Most ideas fail

The ideas that succeed take a long time

ConceptDev

Model Eval Dev

Page 12: Isca needle a_0610

The ideas that succeed take a long time

So aim research 5-10 years ahead of current practice

Page 13: Isca needle a_0610

Current Architecture Practice

Page 14: Isca needle a_0610
Page 15: Isca needle a_0610

5-10 years

Aim Here

Page 16: Isca needle a_0610

5-10 years

Enable this point

Page 17: Isca needle a_0610

Timeline for some ideas

Idea Concept Published Product DT

Stream Processing 1995 1998 2006 11

Virtual Channels 1985 1990 1992 7

Equalized Signaling 1995 1996 2000 5

High-Radix Networks 2002 2005 2008 6

Page 18: Isca needle a_0610

The Performance Equation

ckf

CPINITime

Page 19: Isca needle a_0610

The Research Formula

ROI reward

risk effort

Page 20: Isca needle a_0610

Reward

If you are wildly successful, what difference will it make?

ROI reward

risk effort

Page 21: Isca needle a_0610

Effort

Learn as much as possible with as little work as possible

ROI reward

risk effort

Page 22: Isca needle a_0610

Effort

Do the minimum analysis and experimentation necessary to make a point

ROI reward

risk effort

Page 23: Isca needle a_0610

Real and Artificial Constraints

Real Constraints Artificial Constraints

Laws of physicsFuture semiconductor processesPackaging and thermal limitsFuture applications

Existing ISAExisting OSToday’s benchmarksExisting compilersInfrastructure

Page 24: Isca needle a_0610

Constraining Infrastructure

uArch Idea

Other

uArch

ISA

Compiler

Benchmarks

Binaries

Simulator

Page 25: Isca needle a_0610

Constraining Infrastructure

uArch Idea

Other

uArch

ISA

Compiler

Benchmarks

Binaries

Simulator

Page 26: Isca needle a_0610

Constraining Infrastructure

uArch Idea

Other

uArch

ISA

Compiler

Benchmarks

Binaries

Simulator

Page 27: Isca needle a_0610

The contribution is insight

Not novelty

Not numbers

Page 28: Isca needle a_0610

Research is a hunt for insight

Need to get off the beaten path to find new insights

Page 29: Isca needle a_0610

Road-Kill Research

uArch Idea

Other

uArch

ISA

Compiler

Benchmarks

Binaries

Simulator

Page 30: Isca needle a_0610
Page 31: Isca needle a_0610

Looking here for lost keys

Page 32: Isca needle a_0610

Lost keys here

Looking here

Page 33: Isca needle a_0610

The Academic Advantage

Page 34: Isca needle a_0610

The Academic Advantage

Freedom

Page 35: Isca needle a_0610

The Academic Advantage

Freedom from artificial constraints

Freedom to fail (take risks)

Page 36: Isca needle a_0610

Academic research matched for early stages of the funnel

ConceptDev

Model Eval Dev

Page 37: Isca needle a_0610

Example: ELM

An Ensemble Many Ensembles and memory tiles on a die

37

Page 38: Isca needle a_0610

Example: ELM

Balfour et al., "An Energy-Efficient Processor Architecture for Embedded Systems" CAL, Jan. 2008, pp 29-32.

Page 39: Isca needle a_0610

ELM Infrastructure

uArch Idea

Other

uArch

ISA

Compiler

Benchmarks

Binaries

Simulator

Changed for ELM

Page 40: Isca needle a_0610

The Industrial Advantage

Resources and Experience

Page 41: Isca needle a_0610

The Industrial Advantage

Resources to carry out detailed studies

Experience to address commercial constraints

Page 42: Isca needle a_0610

The ideal partnership:

Academic research 5-10 years out, focused on industry problems

Transfer insight to industrial research to refine into product

ConceptDev

Model Eval Dev

Page 43: Isca needle a_0610

What transfers is insight

Not academic design

Not performance numbers

Page 44: Isca needle a_0610

What transfers is insight

And its transferred by people

Not papers

Page 45: Isca needle a_0610

Concept

Analysis

Simulation

Prototype

Refine Concept

Detailed Design

Academic

Industrial

Page 46: Isca needle a_0610

Concept

Analysis

Simulation

Prototype

Refine Concept

Detailed Design

Academic Industrial

Gap

Paper Impact

Page 47: Isca needle a_0610

Example: Cray T3D and T3E

Page 48: Isca needle a_0610

J-Machine

• MIT 1987-1992

• 3-D network

• Global address space

• Fast messaging and synchronization

• Support for many models of computation

Page 49: Isca needle a_0610

Cray T3D• Started working with Cray in

1989

• Project started early 1990

• First ship in mid 1992

• From J-Machine• Network

• Fast communication/sync

• Global address space

• For reality• Alpha processors

• MECL gate arrays

• Robust software stack

Page 50: Isca needle a_0610

Best Practices for Academics

• Long-term perspective (5-10 years)• Know your customer and their long-term issues

• Look at tomorrow’s applications, not yesterdays

• Maximize reward, minimize effort• Estimate maximum impact – terminate…

• Minimal analysis and experiment to make the point

• Exploit your freedom• Don’t be limited by exiting tools, benchmarks, ISAs, …

• Carry result to impact

• Build relationships with industry

ROI reward

risk effort

uArch Idea

Other

uArch

ISA

Compiler

Benchmarks

Binaries

Simulator

Page 51: Isca needle a_0610

Best Practices for Industry

• Leverage academic research• Build partnerships

• Articulate long-term research issues

• Be open-minded

• Minimize artificial constraints

• Carry concepts across “the gap”

• Open infrastructure

Page 52: Isca needle a_0610

A Partnership

Academe Industry

Filtered, De-risked Concepts

Future issuesInfrastructure

Page 53: Isca needle a_0610

The Startup Path

When you can’t find an appropriate industrial partner, make one.

STAC, Avici, Velio, SPI

Page 54: Isca needle a_0610

Concept

Analysis

Simulation

Prototype

Refine Concept

Detailed Design

Academic

Startup

Page 55: Isca needle a_0610

Startup Pros/Cons

Pros

• Don’t have to convince existing company to change course (until exit)

Cons

• Have to convince investors (repeatedly)

• Have to build a whole company, not just a development team• Finance, sales, marketing, …

• Limited resources

• Impatient capital

Page 56: Isca needle a_0610

Example: SPI

Date Event

Jan 2004 SPI Incorporated

Nov 2004 First round financing

April 2006 Tapeout Storm-1

Oct 2006 First ship of Storm-1

2007 Software, software, software

2008 Customers in production

Sept 2009 Doors close

Page 57: Isca needle a_0610

Much easier to license technology to an existing company

Page 58: Isca needle a_0610

Starting a company to bring a new semiconductor product to market costs $30M (to cash flow positive)

If it’s a programmable processor, its $70M

Investors want a 10x ROI

Need to see a $700M exit to justify a new processor company

Page 59: Isca needle a_0610

The future of computer architecture

Page 60: Isca needle a_0610

The future of computer architecture

• NOW is an ideal time for research to move the needle

• Computers are drastically changing• Pervasive parallelism

• Energy limited

• Bandwidth constrained

• Opportunity to set the MSB of future computers in the next few years

• Requires changing the whole stack

• Requires industry-academe partnership

Page 61: Isca needle a_0610

Energy-Efficient ArchitectureAbstracting Locality

20mm

7pJ

50pJ 500pJ

2000pJ

2000pJ

P P P P

L1 L1 L1 L1

Net

L2

Net

L3

Page 62: Isca needle a_0610

Solution involves many levels of the “stack”

Application

Algorithm

Prog. System

Compiler

ISA

uArch

Design

Circuits

Process

Too constrained to innovate within one layer

Page 63: Isca needle a_0610

Industry

Academe

ROI reward

risk effort

uArch Idea

Other

uArch

ISA

Compiler

Benchmarks

Binaries

Simulator

Page 64: Isca needle a_0610

Moving the NeedleComputer Architecture Research

in Academe and IndustryBill DallyChief Scientist & Sr. VP of Research, NVIDIABell Professor of Engineering, Stanford University