GAP

1
GAP

• Graphical Asymmetric Processing
• SBIR/Final Presentation
• CS410, Group 2
• May 3rd, 2004

2
Project Team Members
3
Problem - Inefficiency

• Universities and Businesses are
• Being inefficient with time
• Being inefficient with money
• Every time they buy a computer

4
Solution- The Graphics Processor

• Why?
• Computers have a GPU
• GPUs are processors, like a CPU
• GPUs are not being used by businesses, even in
  the most stressful tasks.

5
Utilize it!

• A product can be created that will stop this
  problem
• Use the unused GPU
• Improve Return On Investment time
• Improve throughput
• NO extra hardware.

6
Description

• Utilizing existing hardware
• Improve computing power
• Improve computing time
• Improve computing responsiveness
• By creating a
• Programmer API to use the GPU
• Selling that API to software developers

7
Project Goals

• Increase computer performance
• Leverage the GPU (no extra cost)
• Keep compatibility with current software
• Solution should be cross platform

8
Computer Internals
CPU
GPU
9
The Computer Highway

• Data flows to and from the CPU much like cars on
  a highway when there is too much information
  there is a traffic jam which slows down everyone.

10
GPU 200 GigaFLOPS Theoretical (GeForceFX) CPU
170 GigaFlops Theoretical (PowerPC
G5) Before After Usage 75 CPU
efficiency, no GPU Usage CPU _at_ 60, GPU _at_
30 CPU128 GigaFLOPS CPU102
GigaFLOPS GPUgt1 GigaFLOPS GPU51 GigaFLOPS
128 GigaFlops 153 GigaFLOPS (20 Gain)
Source SPEC Shark
11
Supporters

• ATI and nVidia, two major video card
  manufacturers, both provide the technology to do
  this and on multiple platforms.

12
Market Share for Standalone GPUs
Mercury Research Details 2003 Graphics
Market http//www.xbitlabs.com/news/video/display/
20031029073519.html
13
Supporters

• Dr. Varadarajan, head of VA Techs super
  computer, was asked about the ability to leverage
  these processors and agreed it was feasible and
  desirable.

14
GPUs Raw Performance

• Performance growth has multiplied at a rate of
  2.8 times per year since 1993
• Analysts expect this pace to maintain for another
  5 years
• GPU performance will move into the teraflop range
  by 2005

http//www.computer.org/computer/homepage/1003/ent
ertainment
15
Previous Research

• Discussed since 1996, but never implemented.
• Main research issues include quality of floating
  point
• The numbers are single precision not double.
• Works best when batched, which requires a
  relatively parallel system
• Already a multithreading issue. Solutions both in
  programmer practice and compiler design exist.

16
Similar Technologies

• The PowerPC/Macintosh AltiVec processor, AKA
  Velocity Engine
• The Pentium3/Pentium4s SSE/SSE2 instructions,
  AKA NetBurst
• 3dFXs revolution with the Voodoo and Glide in 3d
  graphics in the mid-90s.

17
The Computer Highway (Cont)

• Our solution gives some of the data another path
  that to the graphics card. Functionally it will
  act very much like an HOV lane to being processed.

18
Procedures

• Extreme Programming Programming
• Eight years old, proven
• Strengths
• Communication, Simplicity, Feedback, Courage
• Ideal for dynamic systems and experimental
  projects

FOR MORE INFO...
http//www.extremeprogramming.org/
19
Management Plan
Management Plan
Marketing Plan
Finance Plan
Scheduling Plan
Evaluation Plan
Print Ads
SBIR
Demos
XP
20
The Market

• API implementation sold to developers who use the
  technology to write their programs.
• over 585,000 people are professional developers
  making the 51st most populous job in the nation
  (eighth highest among skilled labor) according
  to the Department of Labor. More are in
  universities and overseas.

21
The Market (cont)

• Would require a recompile of existing software
  and to be optimal probably code modifications
  (particularly for single-threaded applications).
• Case-in-points of new APIs Intel SSE/SSE2, 3dFX
  Glide, Microsoft .NET

22
PC Sales (Millions of Units)
http//www.etforecasts.com/products/ES_pcww1203.ht
m
23
PCs in Use
Computer Industry Almanac http//www.c-i-a.com/p
r0302.htm
24
Server CPUs Have Just Enough Power
 http//www.adaptec.com/worldwide/product/markedit
orial.html
25
Customer Characteristics

• Employs programmers involved in full scale
  development
• Uses applications requiring large amounts of
  computing power
• Has need to analyze data quickly

26
Businesses Identified

• Consumer goods industries
• Retailers
• Financial services industry

27
BI Regulates a Company's Operations

• BI is no longer a luxury it is essential
• Decisions are made at lower and lower levels in
  organizations
• Everyone is an analyst at some point in the day
• Front-line workers are increasingly responsible
  for making large decisions
• They must have access to the results of BI to
  make appropriate decisions

28
OLAP Products Criteria

• Performance
• Calculation times
• Load times
• Multi-user write-back
• What-if capabilities
• Ease of deployment
• Sophisticated modeling
• Fully integrated, enterprise-wide workflow
  collaboration

www.opalreport.com/Applications.htm
29
OLAP Market
The OLAP ReportMarket Share Analysis
http//www.olapreport.com/Market.htmshares
30
Worldwide Analytic Apps SW Market

• Over 4.8 billion in 2007
• CRM analytics 12.9
• Financial analytics 10.3
• Operations analytics 7.4

31
A Few Statistics

• IDC studied analytic applications and their
  impact on core business processes
• What was the ROI for business analytics?
• 46 of organizations generated ROI lt 100
• 34 generated ROI between 101 and 1,000
• 20 generated ROI gt 1,000

The Financial Impact of Business Analytics,
www.idc.com/analyticsroi/
32
Data Mining Market

• Research firm IDC expects 13 compound annual
  growth rate through 2006

http//www.pcmag.com/article2/0,1759,646111,00.asp

33
Case Study The World Bank

• ITs mission fight poverty with passion
  professionalism
• Goal to perform quick, insightful analyses on
  many economic indicators from a hundred countries
  
• How - BI environment that
• Allows users to perform complex calculations in
  seconds in a context and format relevant to
  individual users
• Provides trustworthy economic data analyzed on
  the fly

www.hpcwire.com/dsstar/00/1003/102243.html
34
Case Study The World Bank

• Results.
• Economists have analytical, collaborative
  capabilities contributing to improved oversight -
  making governments more accountable.
• Great variety of skills in users satisfied
  heads of state, finance ministers, chief
  economists, educators, etc.

35
Results from the World Bank

• They gained
• Increased business performance
• Reduced operation wastes
• Improved customer relations

36
Competition Matrix
37
Competitive Analysis

• Competitor
• Cheaper new computers
• Our Strengths against them
• Extendable to current and future machines
  (combinable)
• No extra logistics requirements (power, cooling,
  etc)
• Our Weaknesses against them
• New and untested use of technology
• Method of integration not as clear cut to a
  manager.

Virginia Tech Terrascale Cost Comparison
38
Competitive Analysis, Cont.

• Competitors
• Custom super-computers (i.e. Cray)
• Our Strengths Against Them
• Cheaper
• No Vendor lock-in
• Our Weaknesses Against Them
• Less support
• No reputation

Virginia Tech Terrascale Cost Comparison
39
Selling Method

• Sell to Software Developers
• Unit Price estimation 500/per license/per year
• Estimated Market Over 350 companies (20 of
  high end computing software developers)

40
Versatility in Sales

• Microsoft, Apple, Sun
• Sellable after Proof of Concept stage for fast
  turn around and no need for specialist
  programmers or long term investment
• Individual License
• Sellable later to programmers and/or users for
  possibility of greater profit margins

41
Phase 0
42
Phase 1
43
Staffing Plan Phase 1

• Manager
• 4 Programmers
• Technical Documenter
• Web Developer

44
Phase 1 Budget
 
 
45
Phase 2
46
Staffing Plan Phase 2

• Manager
• 4 Programmers
• Marketing

47
Phase 2 Budget
 
 
48
Phase 2 Scheduling

• A Proof of Concept could be done within a month
  given programmers who know Assembly and HLSL
  (High Level Shader Language). Refinements can
  continue while handing off more complicated (and
  thus costly) programming needs to card
  manufacturers.

49
Phase 2 Scheduling

• Cost could be reduced for GAP by chipset makers
  supporting it through drivers, much like they
  today support OpenGL and Direct3d.
• nVidia, the largest video card manufacturer, has
  more software engineers than hardware engineers.

50
Phase 3
51
Staffing Plan Phase 3

• Manager
• Programmer
• Marketing

52
Phase 3 Budget
53
Phase Deliverables

• Phase 0
• Concept
• Feasibility
• Milestones Identified
• SBIR I Proposal
• Phase 2
• Product Sales
• Driver Model

• Phase 1
• Prototype
• Documentation
• Web Site
• SBIR II Proposal
• Phase 3
• Product to market
• Refine Performance

54
Reasons economic and scientific

• Tools sold to developers have price tags from
  500-50,000 and include performance optimizers.
• Case in Point Maya Unlimited is 7000 USD at
  this date.
• With the growing power of GPUs, it would be a
  missed opportunity to not leverage them.

55
Conclusion

• There is a need for increased computing power
• Our product provides the ability for software
  developers to utilize more power with their
  customer using the same hardware
• Our solution is relatively fast and cheap to
  create
• It can be leveraged across a computers of any
  size huge clusters to home users.