© 2009 IBM Corporation Motivation for HPC Innovation in the Coming Decade Dave Turek VP Deep Computing, IBM

© 2005 IBM Corporation Page 2 2 High Performance Computing Trends  Three distinct phases. Past: Exponential growth in processor performance mostly through CMOS technology advances Near Term: Exponential (or faster) growth in level of parallelism. Long Term: Power cost = System cost ; invention required  Curve is not only indicative of peak performance but also performance/$ Past Near Term Long Term 1PF: PF: EF: 201X?

© 2005 IBM Corporation Page G100G1T10T100T1P10P100P1E Power6 QS22 Blade QS22 Rack BG/P Rack Roadrunner BlueGene/P P505Q Rack 10 BG/P Racks 2 Days forecast10 Days forecastMulti-Scale Multi-Physics Climate Models AstrophysicsOcean modelsGlobal warming Hurricane Models Engineering Geosciences Energy (nuclear) Vibroacustic analysisComp. Photo-Lithogr.Plasma ( Fusion ) Full aircraft design Solid Earth (Petroleum, Water, Voids) Nuclear FissionFull automobile Earthquake Airfoil design Weather Climate Life Sciences In vivo bone anal. Peptide analysisMouse brain Full bone anal. Rat brainProtein Folding Human brainG-receptors Rigid dockingMassive rigid dockingFirst principle docking Free Energy based docking Materials Modeling First Principle device simulations Phase TransitionsComp. Spectroscopy Electron Transfer Electronic Structure Calculations Nano-scale modeling Multi-scale Material SimulationsHigh-k materials P575 Rack

© 2005 IBM Corporation Page 4 4  Core Frequencies ~ 2-4 GHz, will not change significantly as we go forward 100,000,000 Cores to deliver an Exaflop  Power At today’s MegaFlops / Watt: 2 GW needed (~$2B/yr) Power reduction will force simpler chips, longer latencies, more caches, nearest neighbor network  Memory and Memory Bandwidth Much less memory / core (price) Much less bandwidth / core (power / technology)  Network Bandwidth Much less network bandwidth per core (price / core) (Full fat tree ~$1B to $4B) Local network connectivity  Reliability Expect algorithms / applications will have to permit / survive hardware fails.  I/O Bandwidth At 1 Byte / Flop, an EXAFLOP system will have 1 EXABYTE of Memory. No disk system can read / write this amount of data in reasonable time. (BG/P 4TB ~1min but disk array ingest at ~15min) Computer Design Challenges  Exascale Computing O(100 M) compute engines working together  Capability delivered has the potential to be truly revolutionary  However Systems will be complex Software will be complex Applications will be complex Data Centers will be complex Maintenance / Management will be complex

© 2005 IBM Corporation Page 5 5 Summary  Why Exascale? Applications not possible with smaller machines Applications with multiple integrated components for complex systems Applications needing many iterations for sensitivity analysis, etc  Exascale has enormous challenges! Power Cost Memory requirements Usability  Users will need time on a successive series of larger platforms to get to the exascale. Code development will be a large undertaking and tools to assist in this effort are critical. Thank You Capability Understanding Complexity