What s New with GPGPU?

Transcription

1 What s New with GPGPU? John Owens Assistant Professor, Electrical and Computer Engineering Institute for Data Analysis and Visualization University of California, Davis

2 Microprocessor Scaling is Slowing 1e+7 1e+6 1e+5 1e+4 1e+3 1e+2 1e+1 52%/year ps/gate 19% Gates/clock 9% Clocks/inst 18% 19%/year Perf (ps/inst) 1e [courtesy of Bill Dally] [courtesy of Bill Dally]

3 Today s Microprocessors Scalar programming model with no native data parallelism SSE is the exception Few arithmetic units little area Optimized for complex control Optimized for low latency not high bandwidth Result: poor match for many apps Pentium III 28.1M T

4 Future Potential is Large 1e+7 1e+6 1e+5 1e+4 1e+3 1e+2 1e+1 1e+0 1e-1 1e-2 1e-3 52%/year 74%/year 30:1 Perf (ps/inst) Linear (ps/inst) 19%/year 1,000:1 1e ,000:1 2001: 30:1 2011: 1000:1 [courtesy of Bill Dally]

5 Parallel Processing is the Future Major vendors supporting multicore Intel, AMD Excitement about IBM Cell Hardware support for threads Interest in general-purpose programmability on GPUs Universities must teach thinking in parallel

6 Long-Term Trend: CPU vs. GPU GPU

7 Recent GPU Performance Trends Programmable 32-bit FP multiplies per second 54.4 GB/s 49.6 GB/s 6 GB/s $278 X1900 XTX $ GTX $334 P4X3.4 Data courtesy Ian Buck; from Owens et al [EG STAR]

8 Functionality Improves Too! 10 years ago: Graphics done in software 5 years ago: Full graphics pipeline Today: 40x geometry, 13x fill vs.. 5 yrs ago Programmable! Programmable, data parallel processing on every desktop The GPU is the first commercial data- parallel processor

9 The Rendering Pipeline Application Geometry Compute 3D geometry Make calls to graphics API Transform geometry from 3D to 2D (in parallel) Rasterization Generate fragments from 2D geometry (in parallel) Composite GPU Combine fragments into image

10 The Programmable Rendering Pipeline Application Geometry (Vertex) Compute 3D geometry Make calls to graphics API Transform geometry from 3D to 2D; vertex programs Rasterization (Fragment) Generate fragments from 2D geometry; fragment programs Composite GPU Combine fragments into image

11 NVIDIA GeForce D Pipeline Vertex Triangle Setup Z-Cull Shader Instruction Dispatch Fragment L2 Tex Fragment Crossbar Composite Memory Partition Memory Partition Memory Partition Memory Partition Courtesy Nick Triantos, NVIDIA

12 Programming a GPU for Graphics Application specifies geometry rasterized Each fragment is shaded w/ SIMD program Shading can use values from texture memory Image can be used as texture on future passes

13 Programming a GPU for GP Programs Draw a screen-sized quad Run a SIMD program over each fragment Gather is permitted from texture memory Resulting buffer can be treated as texture on next pass

14 GPUs are fast (why?) Characteristics of computation permit efficient hardware implementations High amount of parallelism exploited by graphics hardware High latency tolerance and feed-forward dataflow allow very deep pipelines allow optimization for bandwidth not latency Simple control Restrictive programming model Competition between vendors

15 but GPU programming is hard Must think in graphics metaphors Requires parallel programming (CPU-GPU, task, data, instruction) Restrictive programming models and instruction sets Primitive tools Rapidly changing interfaces Big picture: Every time I say GPU, substitute parallel processor

16 GPGPU in Scientific Visualization Flowfield Vis: a million particles with 60 fps Tensorfield Vis: CPU GPU Speedup = 1 : 150 (!!!) Online decompression: Interactively Visualize 20 GB of data on a 256 MB GPU GPGPU and VIS a winning team Focus and Context Raycasting: High Quality interactive visualization of large datasets on $400 hardware (winner of IEEE 2005 VIS Competition) Courtesy Jens Krüger (TU München)

17 GPGPU Effects Fire Water Let your Virtual Reality come to live with interactive GPGPU effects! (all off these effects are simulated and rendered on the GPU in realtime) Smoke Courtesy Jens Krüger (TU München)

18 Challenge: Programming Systems Programming Model High-Level Abstractions/ Libraries Low-Level Languages Performance Analysis Tools Compilers Docs CPU Scalar STL, GNU SL, MPI, C, Fortran, gcc,, vendor-specific, gdb, vtune,, Purify, Lots applications GPU Stream? Data-Parallel? Brook, Scout, sh, Glift GLSL, Cg, HLSL, Vendor-specific Shadesmith, NVPerfHUD None kernels

19 Glift: Data Structures for GPUs Goal Simplify creation and use of random-access GPU data structures for graphics and GPGPU programming Contributions Abstraction for GPU data structures Glift template library Iterator computation model for GPUs Aaron E. Lefohn, Joe Kniss, Robert Strzodka, Shubhabrata Sengupta, and John D. Owens. Glift: An Abstraction for Generic, Efficient GPU Data Structures '. ACM TOG Jan 2006.

20 Today s Vendor Support High-Level Graphics Language OpenGL D3D Low-Level Device Driver

21 Possible Future Vendor Support High-Level Graphics Language OpenGL D3D Compute Low-Level Device Driver High-Level Compute Lang. Low-Level API

22 ATI CTM Low-level interface to GPU

23 Big Picture Research Targets We don t t know what architecture will win. But we know it will be parallel. than bottom-up what SHOULD we support? Data structures Top-down approach rather Interaction of algorithms and data structures Export to multiple architectures Self-tuning code Communication between GPUs Programming systems for multiple parallel architectures Major obstacle: difficulty of programming. Danger of fragmentation! Opportunity for education as well. Learn from past Explore portable primitives What we re good at: using GPUs as first class computing resources

24 Rob Pike on Languages Conclusion A highly parallel language used by non-experts. Power of notation Good: make it easier to express yourself Better: hide stuff you don't care about Exposing Parallelism Best: hide stuff you do care about Control Flow Give the language a purpose. Data Locality Synchronization

25 Moving Forward What will DX10 give us? What works well now? What doesn t t work well now? What will improve in the future? What will continue to be difficult?

26 The New DX10 Pipeline Vertex Vertex Buffer Buffer Input Input Assembler Assembler Vertex Vertex Shader Shader Setup Setup Rasterizer Rasterizer Output Output Merger Merger Pixel Pixel Shader Shader Geometry Geometry Shader Shader Index Index Buffer Buffer Texture Texture Texture Texture Render Render Target Target Depth Depth Stencil Stencil Texture Texture Stream Stream Buffer Buffer Stream out Stream out Memory Memory memory memory programmable programmable fixed fixed Sampler Sampler Sampler Sampler Sampler Sampler Constant Constant Constant Constant Constant Constant [courtesy David Blythe]

27 What Runs Well on GPUs? GPUs win when Limited data reuse P4 3GHz NV GF 6800 Memory BW 6 GB/s 36 GB/s Cache BW 44 GB/s -- High arithmetic intensity: Defined as math operations per memory op Attacks the memory wall - are all mem ops necessary? Common error: Not comparing against optimized CPU implementation

28 Arithmetic Intensity Historical growth rates (per year): Compute: 71% DRAM bandwidth: 25% DRAM latency: 5% GFLOPS 7x Gap R300 R360 R420 GFloats/sec [courtesy Ian Buck]

29 Arithmetic Intensity GeForce 7800 GTX Pentium GHz GPU wins when Arithmetic intensity Segment 3.7 ops per word 11 GFLOPS [courtesy Ian Buck]

30 Memory Bandwidth GPU wins when Streaming memory bandwidth SAXPY FFT GeForce 7800 GTX Pentium GHz [courtesy Ian Buck]

31 Memory Bandwidth Streaming Memory System Optimized for sequential performance GPU cache is limited Optimized for texture filtering Read-only GeForce 7800 GTX Pentium 4 Small Local storage CPU >> GPU [courtesy Ian Buck]

32 What Will (Hopefully) Improve? Orthogonality Instruction sets Features Tools Stability Interfaces, APIs, libraries, abstractions Necessary as graphics and GPGPU converge!

33 What Won t Change? Rate of progress Precision (64b floating point?) Parallelism Won t sacrifice performance Difficulty of programming parallel hardware but APIs and libraries may help Concentration on entertainment apps

34 GPGPU Top Ten The Killer App Programming models and tools GPU in tomorrow s computer? Data conditionals Relationship to other parallel hw/sw Managing rapid change in hw/sw (roadmaps) Performance evaluation and cliffs Philosophy of faults and lack of precision Broader toolbox for computation / data structures Wedding graphics and GPGPU techniques