Grice-QS22

2008 IBM CorporationIBM CONFIDENTIAL

2008 IBM Corporation

IBM Systems and Technology Group

Cell/B.E. processor-based systems and software offeringsIBM BladeCenter QS22 and SDK 3.0


2008 IBM Corporation2 Sales Conference

The challenge today

For many years, organizations have relied on performance gains from increasing clock speeds of traditional microprocessor architectures

This approach has been challenged by the physical limitations ofsemiconductors and by traditional processor architecture implementations

High performance computing (HPC) applications need a fundamentally new technology and approach to the system-level architecture to achieve the desired level of performance.



Cell Broadband Engine (Cell/B.E.) Technology

IBM, Sony, Toshiba Alliance formed in 2000

March, 2001 STI Design Center opened in Austin, TX

April, 2004 - Single Cell BE operational

July, 2004 - 2-way SMP operational

February, 2005 - first technical disclosures at ISSCC

May, 2005 - first public demonstration of Cell/B.E. processor-based system at E3

August, 2005 - published technical details of Cell/B.E. architecture

November, 2005 - published open source SDK & Cell/B.E. simulator

August, 2006 - introduced the very first Cell/B.E. processor-based server to the market

For a higher of absolute performance and efficiency



IBM commitment to innovation

2006

2008

2007Produce systems for early adoption and solution enablement

Create initial platforms for experimentation

BladeCenter QS21

IBM SDK forMulticoreAcceleration 3.0

BladeCenter QS20Produce robust production ready systems for targeted industry applications

IBM BladeCenter QS22Extraordinary double precision floating point performance. Large memory capability. Ready for the most demanding production applicationsPowerXCell 8i processor



Cell Broadband Engine Architecture (CBEA) Technology Roadmap

20102009200820072006

PerformanceEnhancements/Scaling

CostReduction

All future dates and specifications are estimations only; Subject to change without notice. Dashed outlines indicate concept designs.

ConceptCommitted

Compatible code and security base across entire lineCompatible code and security base across entire line

Cell/B.E.(1+8)

90nm SOI

IBM PowerXCell 8i(1+8eDP SPE)

65nm SOI

Cell/B.E.(1+8)

65nm SOI

IBM PowerXCell 32ii

45nm SOI

Cell/B.E.(1+8)

45nm SOI



IBM PowerXCell 8i processor benefits

Sets a new performance standard Accelerates computationally intense workloads such as

analytics, multimedia and vector processing. Efficient computation per watt

Designed for flexibility Wide variety of application domains Cell can cover a wide range of application space with its

capabilities in floating point operations, integer operations data streaming / throughput support real-time support

Exploits C/C++, Fortran programming models Enhanced security capability

Virtual trusted computing environment for security

The new PowerXCell 8i processor builds on the Cell Broadband Engine Architecture and combines a general-purpose Power Architecture core of modest performance with eight enhanced synergistic processing elements optimized for extreme double precision and single precision computational performance

PowerXCell 8i processor 65 nm 9 cores, 10 threads 230.4 GFlops peak (SP) at 3.2GHz 108.8 GFlops peak (DP) at 3.2GHz Up to 25 GB/s memory bandwidth Up to 75 GB/s I/O bandwidth 92 Watts @ 3.2GHz Top frequency >4GHz

(observed in lab)

PowerXCell 8i processor 65 nm 9 cores, 10 threads 230.4 GFlops peak (SP) at 3.2GHz 108.8 GFlops peak (DP) at 3.2GHz Up to 25 GB/s memory bandwidth Up to 75 GB/s I/O bandwidth 92 Watts @ 3.2GHz Top frequency >4GHz

(observed in lab)



Intels x86 Quad Core processors are Dual Chip Modules (DCMs), 2 of these processor

stacked vertically & packaged together

PowerXCell 8i uses the space & power and delivers more than 2.3x the GFlops of traditional architecture

On any traditional processor, shown ratio of cores to cache, prediction, & related items

illustrated here remains at ~50% of area the chip area

Example Server Dual Core

349mm2, 3.4 GHz @ 150W2 Cores, ~27.2 SP GFlops1.3b Transistors @ 65nm

Example Desktop Quad Core

214 mm, 3 GHz @ 130W4 Cores, ~96 SP GFlops

820m Transistors @ 45nm

PowerXCell 8i Nine Core

109 mm2 3.2 GHz@ 75W9 cores, ~ 230 SP GFlops,250m Transistors @ 65nm



BladeCenter QS22 PowerXCell 8i Core Electronics

Two 3.2GHz PowerXCell 8i Processors SP: 460 GFlops peak per blade DP: 217 GFlops peak per blade Up to 32GB DDR2 800MHz Standard blade form factor Support BladeCenter H chassis

Integrated features Dual 1Gb Ethernet (BCM5704) Serial/Console port, 4x USB on PCI

Optional Pair 1GB DDR2 VLP DIMMs as I/O buffer

(2GB total) (46C0501) 4x SDR InfiniBand adapter (32R1760) SAS expansion card (39Y9190) 8GB Flash Drive (43W3934)

DDR2

DDR2

DDR2

DDR2

PowerXCell 8i

DDR2

PowerXCell 8i

2 UART, SPI

Rambus FlexIO

PCI-E x16PCI-X

PCI-E x8

HSC *12x PCI-E

x16

PCI

Leg

acy C

on

USB toBC mid plane

GbE toBC mid plane

2x1GbE

SPI

Optional IB2 port

IB x4 HCA

HSDC

IB-4x toBC-H high speed fabric/mid plane

DDR2

DDR2

DDR2

DDR2

FlashDrive

DDR2

IBMSouthBridge

4xUSB2.0

Flash, RTC& NVRAM

IBMSouthBridge

DDR2

*The HSC interface is not enabled on the standard products. This interface can be enabled on customsystem implementations for clients by working with the Cell services organization in IBM Industry Systems.



Performance highlights

Performance is an order of magnitude better than general purposeprocessors (GPP) for media and certain applications that can take advantage of its Single Instruction Multiple Data (SIMD) capability

Performance of its simple Power Processor Element (PPE) is comparable to a traditional GPP performance

Each Synergetic Processor Element (SPE) is able to perform mostly the same as a GPP running at the same frequency

Key performance advantage comes from its eight de-coupled SPE engines with dedicated resources including large register files and DMA channels

Accelerates targeted applications with extraordinary processing capabilities Floating-point operations Integer operations Data streaming / throughput support Real-time support

Open architecture allows for optimization at compiler and application level Performance gains from tuning compilers and applications can be significant Tools/simulators are provided to assist in performance optimization efforts



IBM BladeCenter QS22

QS22 is the RIGHT choice for intensive streaming and/or single and double precision floating point workloads

QS22 is OPEN based on Power Architecture and running Linux OS

QS22 is EASY to deploy and to integrate into the existing IT infrastructure and/or workloads: Co-exist and complement all other Blade servers offerings (Intel, AMD, POWER) Ready to scale out and deploy in production environments

QS22 is GREEN more than 1.7 SP (or 0.8 DP) GFLOPS per watt.

Premier blade for HPC workloads



IBM SDK for Multicore Acceleration and related tools

Libraries and frameworks

IBM XL C/C++ compiler*Optimized compiler for use in creating Cell/B.E. optimized applications. Offers:

* improved performance * automatic overlay support * SPE code generation

AcceleratedLibrary

Framework (ALF)

DataCommunication

andSynchronization

(DaCS)

Basic LinearAlgebra

Subroutines (BLAS)

StandardizedSIMD math

libraries

GNU tool chainPerformance

Tools

The IBM SDK is a complete tools package that simplifies programming for the Cell Broadband Engine Architecture

XLC compiler is a

complementary product to SDK

Eclipse-based IDE

Simulator

Denotes software components included in the SDK for Multicore Acceleration



IBM SDK for Multicore Acceleration value

Designed to be highly reliable, simple toacquire and easy to use

Complete, integrated kit Production-ready tools from IBM IBM warranty and support

Based on industry standards to ease thetransition to the Cell/B.E.

Eclipse-based Integrated DevelopmentEnvironment

Standard, base libraries Third-party libraries can be plugged in

Designed to make it easy to port and optimize applications for the QS21 and QS22

Enhancements to enable new features in QS22 Performance tuning tools to help optimize algorithms without re-writing the entire application Tools designed to help you partition an application across a hybrid Cell/B.E. and x86 platform



Cell Programming Approaches are fully customizable!

3. Case Tools / CompleteHardwareAbstraction

User tool-driven

2. AssistedProgramming

Libraries,Frameworks

1. NativeProgramming

Compilers,Intrinsics,DMA, etc.

Increasing Programmer

Control over

Cell/B.E

. resources

Decreasing programmer

attention to

architectural details



Workloads ideal for PowerXCell 8i and QS22

DigitalMedia

Financial ServicesSector

Home MediaConsumer Electronics

Information Based

Medicine

Digital Video Surveillance

Aerospace and Defense

Electronic Design

Automation

Chemicals & Petroleum

Market & Solution Specific Assets

Real-time AnalyticsProcessing of Data

Information SynthesisAnalysis

Unstructured DataMultimodal SearchData TransformsPattern Matching

Image/Video Creation/MgtPresentation of Data

VisualizationImaging

Extreme Stream Computation and Bandwidth requirements

PowerXCell 8i is suited for applications which demand extraordinary floating point performance



Public sector HPC solutions

IBM components: IBM BladeCenter QS21 & QS22 IBM SDK for Multicore Acceleration IBM Cell/B.E. math libraries IBM hybrid computing solution (custom offering) PXCAB

ISV applications: Development tools from RapidMind, Gedae,

Wind River, etc. A growing number of university and government

research labs with external collaborative missions are exercising existing and emerging science codes

The solution is designed to offer: Petaflop Scalability and reliability Lower power and space footprint Lower total cost of ownership

Performance advantages: Science code such as SPaSM, VPIC, Milagro,

Sweep3D, accelerated up to 4-9X faster than AMD Opteron single core(Source: LANL - www.lanl.gov/roadrunner)

Enable government labs, agencies, and academic research centers to run high performance codes faster, less expensively, and with lower power consumption than existing computing architectures

*See Notes on Benchmarks, charts 46 and 47



Aerospace & defense solutions


ISV applications: Gedae stream, image and signal programming

environment RapidMind development tools Wind River VxWorks RTOS and WorkBench

Tools

Performance advantages: FFT workloads up to 7.7x faster than 3.0 GHz

2-core Woodcrest x2* Double Precision Matrix Multiplication up to

2.6x faster than 2.66GHz 4-core Clovertown*

Enhance competitiveness, demonstrate innovation and capture significant government contracts through dramatic performance improvements in real time signal and image processing

As a time-served radar architect, I can say that Cell/Gedae is something of a dream and should

rightly impact the new design market it is an opportunity that the DoD should not fail to grasp.

- John Roulston,SCImus Solutions, March 2007




Digital content creation solutions

IBM components: IBM BladeCenter QS21 & QS22 IBM SDK for Multicore Acceleration IBM Cell/B.E. math libraries IBM hybrid computing solution (custom offering) PXCAB IBM iRT scalable real-time ray tracer

ISV applications RapidMind development tools

The solution is designed to offer: Rapid turn around of digital assets More realistic simulation An open and flexible solution based on standards Scalability and reliability

Performance advantages: 1080p Ray-traced images computed in

milliseconds* 1080p Ambient Occlusion images computed in

seconds*

IBM solutions enable Media and Entertainment companies to produce the next generation of animated feature films, games, and advertising content




Digital video surveillance solutions

IBM components: IBM BladeCenter QS21/QS22 IBM Total Storage IBM DVS ADK

ISV applications: Codec libraries Video distribution software

The solution is designed to offer: H.264 encoding Encoders for analog cameras Transcoding to save storage and

network costs Decoding acceleration to reduce

workstation costs and improve robustness Better management and scalability Network-based surveillance Compute density - with two processors per

blade, 14 blades to a chassis, and two chassis to a rack, it is possible to have as many as 672 H.264 encoders in the rack

Performance advantage: One Cell/B.E processor running at 3.2 GHz,

can encode 12 channels of standard definition video at 30 fps to H.264 (main profile, including CABAC)[1][1] Source: IBM Research benchmark

Solutions deliver hardware and enablement for high-density, highly scalable encoding, transcoding, and compositing for digital video surveillance

IBM BladeCenter QS21/QS22

PTZ

Coax16 camera inputs

16 camera inputs

Aggregation Unit

14 card slotsIBM BladeCenter-HIBM Total Storage

672 encoders in

a rack!




EDA solutions

IBM components: Cell/B.E. hybrid cluster IBM BladeCenter QS21 IBM System x / IBM BladeCenter IBM Cluster 1350 integrated cluster Storage: DS4000, N series, DCS9550

ISV applications: Mentor Graphics Calibre nmOPC

and OPCVerify

The solution is designed to offer: Significant run time acceleration

Leverages Cell/B.E. strengths to offer significant speed-up when compared to existing solutions in the market, reducing design turnaround time

Scalability and reliability Blade form factor improves scalability,

compute density and reliability

Accelerate computational lithography workload to address turnaround time challenges and at the same time reduce total cost of the computing infrastructure



Financial market analytics solutions

IBM components: IBM BladeCenter QS22 IBM SDK for Multicore Acceleration Dynamic Application Virtualization Cell/B.E. math libraries

ISV applications: NAG - Math & Stat Software Platform Symphony -Grid Computing

Environment Encirq Event Processing Platform

The solution is designed to offer: Flexibility and Scalability

IBM Bladecenter QS22 integrates with other Bladecenter Products

IBM SDK, DAV, third party applications for ease of adoption within existing infrastructure

Technical Services with skilled programming expertise and subject matter experts

Power, space and cooling advantages

Performance advantage Collateralized Debt Obligation (CDO) - 7.5X faster than

2.8 GHz 4-core Harpertown* 650 million European options /sec using Monte Carlo

simulations on QS22 blade*

Enable financial market professionals to perform the required speed, accuracy and highly complex analytics to support trade execution and improve their firms competitive position




Medical imaging solutions


ISV applications: Advanced image and text analytics High-performance image compression

The solution is designed to offer: 3D image reconstruction, registration, volume

rendering, segmentation On-demand compression/decompression

Performance advantage: 16x improvement on MRI image reconstruction

over Opteron system 11x improvement on CT image reconstruction

over 3.0GHz Xeon system 48x improvement on image registration over

3GHz Pentium 4 200x shear-warp volume visualization over TI

TMS320C80 processor 40:1 CT study data compression

(Source for all above: Mayo Clinic -http://www.mayoclinic.org/news2007-rst/3996.html )*

Improve the efficiency, productivity, and quality of patient care through dramatic performance improvements in the transmission and analysis of medical images




Seismic solutions

IBM components: IBM BladeCenter QS22 IBM SDK for Multicore Acceleration IBM Cell/B.E. math libraries IBM hybrid computing solution (custom offering) PXCAB

Standard math, vector math, FFT, BLAS, MPI and tridiagonal solver

ISV applications: Simudyne Customers own proprietary code

The solution is designed to offer: High-performance highly accurate rendering

of geologic structures Cost effective HPC environment that has

significant performance increases Scalability and reliability

Performance advantages: FFT workloads up to 7.7x faster than 3.0

GHz 2-core Woodcrest x2* Double Precision Matrix Multiplication up to

2.6x faster than 2.66GHz 4-core Clovertown*

Improve the speed and accuracy of geologic visualization to reduce the cost of evaluatingpotential targets for oil and gas yielding potential




QS22 summary

The QS22 is based on the new PowerXCell 8i processor built on an enhanced version of the Cell Broadband Engine Architecture

The QS22 offers the capabilities you need for your most demanding computational requirements

Offers extraordinary double precision and single precision floating point performance

Supports up to 32GB of processor memory

IBM is working with ISVs and customers to accelerate workloadson the QS22 in targeted application areas

The QS22 is extremely efficient, offering more than 1.7 SP (or 0.8 DP) GFLOPS per watt of energy

BladeCenter QS22 is Right, Open, Easy and Green

Premier blade for HPC workloads



IBM SDK for Multicore Acceleration summary

Designed to be highly reliable, simple to acquire and easy to use Complete, integrated kit Production-ready tools from IBM IBM warranty and support

RHEL 5.2 Enterprise support

Based on industry standards to ease the transition to the Cell/B.E. architecture

Eclipse-based Integrated Development Environment Standard, base libraries Third-party libraries can be plugged in

Designed to make it easy to port and optimize applications for the QS22 Performance tuning tools to help optimize algorithms without re-writing the entire application Tools designed to help you partition an application across a hybrid Cell/B.E. and x86 platform



Cell/B.E. architecture reaches wide and deep from consumer products to high performance computing

SCE PS3(Cell/B.E. + GPU)

IBM BladeServer(2 Cell/B.E. orPowerXCell 8i)

Roadrunner(16,000

PowerXCell 8i. + AMD)

Mercury 1u Dual CellSony Cell/B.E. Computing Unit

(Cell/B.E. + GPU + AV I/O)

Consumer Business

High Performance ComputingEnterprise

PowerXCell 8i PCI card

(Cell/B.E. + Host)

Common OSs, Infrastructure, Tools, Libraries, Codethe SAME SPE code runs from end to end

Toshiba SpursEngine

(SPUs. + Host)

Mini-Roadrunner

Custom

Increas

ing supp

ort for s

cale

and

datacen

ter

Increas

ing supp

ort for s

cale

and

datacen

ter

Documents

Grice-QS22