Super ComputerSuper Computer

-----------------IBM Deep Blue

Presented by: Yang Jing WenYe TaoLi Fei Nan

A Computer defeated human!A Computer defeated human!

On May 11, 1997, the machine won a six-game match by two wins to one with three draw against world champion Garry Kasparov

Super ComputerSuper Computer

� Definition(from wikipedia): A supercomputer is a computer at the frontline of current processing capacity, particularly speed of calculation.

� Supercomputers are used for highly calculation-intensive tasks

� Tasks like: weather forecasting, climate research, molecular modeling and physicssimulations.

CrayCray--11

AbstractAbstract

� A chess search engine

� A massively parallel system with multiple levels of parallelism

� A strong emphasis on search extensions

� A complex evaluation function

� Effective use of a Grandmaster game database.

IntroductionIntroduction

� ChipTest and Deep Thought500,000 positions/s(ChipTest)700,000 positions/s(Deep Thought)

� Deep Thought 21.Medium-scale multiprocessing.2. Enhanced evaluation hardware.3. Improved search software.4. Extended book� Deep Blue I(Overall Search speed 50-100million positions/s)

� Deep Blue II

Deep Blue IIDeep Blue II

� Redesigned chess chip

1.Completely redesigned evaluation function

2.Added hardware repetition detection

3.Specialized move generation modes

� Double the number of chess chip

� Development of software tools to aid in debugging and match preparation

The summary of systemThe summary of system

System overviewSystem overview

� Deep blue is organized in three layers

1.One processor as master

2.The remainder as workers

3.Chess chip

The Chess Chip The Chess Chip

� Move generation

� evaluation function

� Search control

� Extendability(provide access to a more complicated search control, and additional terms for the evaluation function)

Software SearchSoftware Search

dual credit with delayed extensions

1.Extend forcing/forced pairs of moves

2.Forced moves are expectation dependent

3.Fractional extensions

4.Delayed extensions

5.Dual credit

Extend forcing/forced pairs of movesExtend forcing/forced pairs of moves

� White has an unstoppable winning threat. Black has numerous delaying moves (e.g., checks, mate threats, attacks on high valued pieces, etc.)

� White has a series of sacrifices and immediate threats, which eventually result in checkmate or the win of material greater than that sacrificed.

Forced moves are expectation Forced moves are expectation

dependentdependent

� A move may be forced for one level of expectation and not forced for another. A move that is “fail low”, i.e., below the current level of expectation, is never considered forced in the Deep Blue search.

Fractional extensionsFractional extensions

� It is not feasible to fully extend all the ffp’s without the search exploding.

� One method of addressing this problem is to allow fractional extensions, where an ffp does not get a full 2-ply extension, but rather some smaller amount, say 1.75 ply.

Delayed extensionsDelayed extensions

� Often an isolated ffp is meaningless

� allow ffp’s to accumulate “credit”, and only when sufficient credit is available can it be “cashed in” for an extension.

� By setting this threshold appropriately, extensions are delayed until multiple ffp’s occur in a given path.

Dual creditDual credit

� Separate and accumulate the credit for the two sides separately.

� If either side accumulates sufficient credit to cash in for an extension, the other side must give up an equal amount of credit.

Hardware SearchHardware Search

• take place on the chess chip

• fast but simple

• host processor

� Once a hardware search is initiated, the host processor controlling that chip is free to do other work, including performing the software search and initiating hardware searches on other chips. The host polls the chips to determine when a hardware search has completed. The host can abort a hardware search if needed, e.g., if the search is taking too long, or is no longer relevant.

PARALLEL SEARCHINGPARALLEL SEARCHING

Processor hierarchyProcessor hierarchy

Parallel search implementationParallel search implementation

1. Load balancing

2. Master overload

3. Sharing between nodes

Parallel search performanceParallel search performance

EVALUATION EVALUATION FUNCTIONFUNCTION

OverviewOverview

1. Slow

2. Fast

`̀