1. Reconfigurable Supercomputing 2004

2. (2) Key Issues in HPC  Leveling off of performance Traditional Scalar/Vector – long product cycles, too few vendors MPP Clusters – only marginal performance gains from Intel, AMD, etc., roadmaps  Scalar/Vector vs. Cluster tradeoff Need for dedicated hardware operations Need for price/performance  Dedicated hardware takes too long to develop (ASIC design & build cycles)  Clusters can change frequently, but not substantively  Leveling off of performance Traditional Scalar/Vector – long product cycles, too few vendors MPP Clusters – only marginal performance gains from Intel, AMD, etc., roadmaps  Scalar/Vector vs. Cluster tradeoff Need for dedicated hardware operations Need for price/performance  Dedicated hardware takes too long to develop (ASIC design & build cycles)  Clusters can change frequently, but not substantively

3. (3) HPTi Factors for New Solution  Reconfigurable Computing Introduces dedicated hardware capability… …that can change at speed of thought Nallatech BenNUEY (e.g., PCI-4E: DIME II) Xilinx Virtex family (e.g., BenPRO, BenBLUE)  Faster HPC Development Cycles Algorithms Computational science Feedback to hardware development  Systems integration and management workflow and software  Evolve to Grid computing and Collaborative research/engineering  Reconfigurable Computing Introduces dedicated hardware capability… …that can change at speed of thought Nallatech BenNUEY (e.g., PCI-4E: DIME II) Xilinx Virtex family (e.g., BenPRO, BenBLUE)  Faster HPC Development Cycles Algorithms Computational science Feedback to hardware development  Systems integration and management workflow and software  Evolve to Grid computing and Collaborative research/engineering Better Price/ Performance Accelerate Science… Enterprise Capability …Even Faster Acceleration

4. (4) Delivered and Benchmarked  48 nodes  2u, back-to-back (net 1u/node)  96 FPGA’s  Annapolis Micro  Xilinx Virtex II  34 Tera-Ops  In use today  All Commodity Parts  48 nodes  2u, back-to-back (net 1u/node)  96 FPGA’s  Annapolis Micro  Xilinx Virtex II  34 Tera-Ops  In use today  All Commodity Parts

5. (5) Future: Grid & Collaborative Research  Grid Networked FPGA clusters - multi-tier/multi-peer HPC Intra-cluster flexibility u Processing u FPGA-based high-speed Interconnect u Graphics Intra-board near-real-time reconfigurability Apply Information Environment solution for metacomputing/metaqueueing capability Apply GTK, OGSA to facilitate development, operations  Collaborative Research Distributed visualization, model walkthrough Heterogeneous architectures Researcher independence from sites, specific platforms Model and Data repositories  Grid Networked FPGA clusters - multi-tier/multi-peer HPC Intra-cluster flexibility u Processing u FPGA-based high-speed Interconnect u Graphics Intra-board near-real-time reconfigurability Apply Information Environment solution for metacomputing/metaqueueing capability Apply GTK, OGSA to facilitate development, operations  Collaborative Research Distributed visualization, model walkthrough Heterogeneous architectures Researcher independence from sites, specific platforms Model and Data repositories

6. (6) HPTi Factors for New Solution  Reconfigurable Computing Introduces dedicated hardware capability… …that can change at speed of thought Nallatech BenNUEY (e.g., PCI-4E: DIME II) Xilinx Virtex family (e.g., BenPRO, BenBLUE)  Faster HPC Development Cycles Algorithms Computational science Feedback to hardware development  Systems integration and management workflow and software  Evolve to Grid computing and Collaborative research/engineering  Reconfigurable Computing Introduces dedicated hardware capability… …that can change at speed of thought Nallatech BenNUEY (e.g., PCI-4E: DIME II) Xilinx Virtex family (e.g., BenPRO, BenBLUE)  Faster HPC Development Cycles Algorithms Computational science Feedback to hardware development  Systems integration and management workflow and software  Evolve to Grid computing and Collaborative research/engineering Better Price/ Performance Accelerate Science… Enterprise Capability …Even Faster Acceleration