1. The End of CMOS Scaling will be Good for Space Computing Fault Tolerant Spaceborne Computing Employing New Technologies May 29, 2008 Sandia National Laboratories Erik DeBenedictis (Sandia) Sandia is a multiprogram laboratory operated by Sandia Corporation, a Lockheed Martin Company, for the United States Department of Energy's National Nuclear Security Administration under contract DE-AC04-94AL85000.

2. Overview HPC Parallel Embedded Low Power COTS/Desktop Development $ Productivity Tools Future Low Power Parallel Heterogeneous Productivity Tools Space Computing Rad-Hard ?

3. Clock Rate Flat Lined Clock rate flat lined a couple years ago, as vendors put excess resources into multiple cores This is a historical fact and evident to everybody, so there is little reason to comment on the cause However, it has profound architectural consequences (later slide) 199020102005 Year  100 MHz 1 GHz 10 GHz 2 GHz 4 GHz

4. ITRS Process Integration Spreadsheet Big Spreadsheet –Columns are years –Rows are 100+ transistor parameters –Manual entry of process parameters by year –Excel computes operating parameters –Extra degrees of freedom go to making Moore’s Law smooth – not the best computers

5. Energy (log scale) for Technology created in Government Fab Year Moore’s Law kT 100kT kT Limit Moderates Optimism for Perpetual Exponential Growth 2008  

6. ITRS 2008 Update – April, Konigswinter, Germany International Technology Roadmap for Semiconductors 2008 ITRS Update ORTC [ Konigswinter Germany ITRS ITWG Plenary] A.Allan, Rev 2, [notes on IRC/CTSG More Moore, More than Moore, Beyond CMOS 04/04/08] Industry’s Plans

8. The Architecture Game This is my diagram from a paper to illustrate CMOS architecture in light of CMOS scaling limits [Discuss] 100% CPU Efficiency (can’t do better) Comm ercial Speed Target 100% 50% 25% 12% 6% 3% 19802010200019902020 Year  log(throughput) Power effici- ency Next Moves:  Switch to Vector Arch.  Switch to SIMD Arch.  Add Coprocessor  Scale Linewidth  Increase Parallelism  Increase Cache  More Superscalar  Raise Vdd and Clk Next Moves Finish

9. 1 2 200820092010 Year  Performance  B Traditional  P with big budget  P with big budget but clock rate and power handicap A Better Idea but with a small budget Special Architectures Go Mainstream Conclusions –Mainstream and embedded technology will become more similar Power Parallelism –Architectures will become more special purpose General systems may be comprised of multiple special purpose sections

10. EXOCHI: Architecture and Programming Environment for A Heterogeneous Multi-core Multithreaded System Perry H. Wang 1, Jamison D. Collins 1, Gautham N. Chinya 1, Hong Jiang 2, Xinmin Tian 3, Milind Girkar 3, Nick Y. Yang 2, Guei-Yuan Lueh 2, and Hong Wang 1 Microarchitecture Research Lab, Microprocessor Technology Labs, Intel Corporation 1 Graphics Architecture, Chipset Group, Intel Corporation 2 Intel Compiler Lab, Software Solutions Group, Intel Corporation 3

11. 11 Future mainstream microprocessors will likely integrate heterogeneous cores How will we program them? Motivation OS My IA CPU My Accelerator Scheduler Process Thread My App Driver Stub Driver API Dispatch My Device Driver ia cpu ia cpu ia cpu ia cpu Map computation to driver / abstraction API Unfamiliar development / debugging flow OS / driver overheads Accelerator in distinct memory space The following 5 Viewgraphs sent by Jamison Collins with permission to post

12. 12 CHI Programming Environment Compiler Modified front-end and OpenMP pragmas –Fork/join –Producer/consumer parallelism Generates fat binary CHI runtime Multi-shredding: User-level threading Extensible to multiple types of heterogeneous cores –E.g. Intel GMA X3000 –E.g. A data streaming systolic array accelerator for communication #pragma omp _asm { …… } Intel C++ Compiler Accelerator-specific assembler and domain-specific plug-ins.code.data.special_section Linker CHI runtime library #pragma omp parallel target(targetISA) [clause[[,]clause]…] structured-block Where clause can be any of the following: firstprivate(variable-list) private(variable-list) shared(variable-ptr-list) descriptor(descriptor-ptr-list) num_threads(integer-expression) master_nowait

13. 13 IA Look-n-Feel: Development and Debugging

14. 14 IA Look-n-Feel: Compilation and Execution

15. Interconnect options CPU Part GPU Part Verilog/ VHDL CPU: 1-core, multi-core FPGAAccelerator, GPU, SIMD, or ASIC Bus/Stream/Message Standards I/O Memory: DRAM, Nano Mass Storage Inter-subsystem gateway RAD- 750, etc. Fault-Tolerant High-Capability Computational Subsystem Spacecraft Control Subsystem Spaceborne Computing with Emerging Technologies Motivation –Greater quantities of data: perform more onboard computing, reduce communications requirements Vision –Multiple computing technologies each used to best advantage Harness advances in semiconductors and nanotech –Need hardware interoperability –Need software tools to support heterogeneous hardware Workshop –Target date May 28-30, 2008 –At Sandia, in and out –Immediate target: Inventory resources and set plans for coordination and standards –Rad hard processing Archival, Maintainable, Source Code