Computing with GPGPUs
Raj Singh
National Center for Microscopy and Imaging Research
GPGPUs and CUDA
Guest Lecture, CSE167, Fall 2008
Graphics Processing Unit (GPU)
Development driven by the multi-billion dollar game industry
Bigger than Hollywood
Need for physics, AI and complex lighting models Impressive Flops / dollar performance
Hardware has to be affordable
Evolution speed surpasses Moores law
Performance doubling approximately 6 months
GPGPUs and CUDA
Guest Lecture, CSE167, Fall 2008
GPU evolution curve
*Courtesy: Nvidia Corporation GPGPUs and CUDA Guest Lecture, CSE167, Fall 2008
GPGPUs (General Purpose GPUs)
A natural evolution of GPUs to support a wider range of applications Widely accepted by the scientific community Cheap high-performance GPGPUs are now available
Its possible to buy a $500 card which can provide almost 2 TFlops of computing.
GPGPUs and CUDA
Guest Lecture, CSE167, Fall 2008
Teraflop computing
Supercomputers are still rated in Teraflops
Expensive and power hungry Not exclusive and have to be shared by several organizations Custom built in several cases
National Center for Atmospheric Research, Boulder installed a 12 Tflop supercomputer in 2007
GPGPUs and CUDA
Guest Lecture, CSE167, Fall 2008
What does it mean for the scientist ?
Desktop supercomputers are possible Energy efficient
Approx 200 Watts / Teraflop
Turnaround time can be cut down by magnitudes.
Simulations/Jobs can take several days
GPGPUs and CUDA
Guest Lecture, CSE167, Fall 2008
GPU hardware
Highly parallel architecture
Akin to SIMD
Designed initially for efficient matrix operations and pixel manipulations pipelines Computing core is lot simpler
No memory management support 64-bit native cores Little or no cache Double precision support.
GPGPUs and CUDA Guest Lecture, CSE167, Fall 2008
Multi-core Horsepower
Latest Nvidia card has 480 cores for simultaneous processing Very high memory bandwidth
> 100 GBytes / sec and increasing
Perfect for embarrassingly parallel compute intensive problems Clusters of GPGPUs available in GreenLight
Figures courtesy: Nvidia programming guide 2.0 GPGPUs and CUDA Guest Lecture, CSE167, Fall 2008
CPU v/s GPU
GPGPUs and CUDA
Guest Lecture, CSE167, Fall 2008
Programming model
The GPU is seen as a compute device to execute a portion of an application that
Has to be executed many times Can be isolated as a function Works independently on different data
Such a function can be compiled to run on the device. The resulting program is called a Kernel
C like language helps in porting existing code.
Copies of kernel execute simultaneously as threads.
GPGPUs and CUDA
Figure courtesy: Nvidia programming guide 2.0 Guest Lecture, CSE167, Fall 2008
Look Ma no cache ..
Cache is expensive By running thousands of fast-switching light threads large memory latency can be masked Context switching of threads is handled by CUDA
Users have little control, only synchronization
GPGPUs and CUDA Guest Lecture, CSE167, Fall 2008
CUDA / OpenCL
A non-OpenGL oriented API to program the GPUs Compiler and tools allow porting of existing C code fairly rapidly Libraries for common math functions like trigonometric, pow(), exp() Provides support for general DRAM memory addressing
Scatter / gather operations
GPGPUs and CUDA
Guest Lecture, CSE167, Fall 2008
What do we do at NCMIR / CALIT2 ?
Research on large data visualization, optical networks and distributed system. Collaborate with Earth sciences, Neuroscience, Gene research, Movie industry Large projects funded by NSF / NIH
NSF EarthScope
GPGPUs and CUDA
Guest Lecture, CSE167, Fall 2008
Electron and Light Microscopes
GPGPUs and CUDA
Guest Lecture, CSE167, Fall 2008
Cluster Driven High-Resolution displays data end-points
GPGPUs and CUDA
Guest Lecture, CSE167, Fall 2008
Electron Tomography
Used for constructing 3D view of a thin biological samples Sample is rotated around an axis and images are acquired for each tilt angle Electron tomography enables high resolution views of cellular and neuronal structures. 3D reconstruction is a complex problem due to high noise to signal ratio, curvilinear electron path, sample deformation, scattering, magnetic lens aberrations
GPGPUs and CUDA
Biological sample
Curvilinear electron path
Tilt series images
Guest Lecture, CSE167, Fall 2008
Challenges
Use a Bundle Adjustment procedure to correct for curvilinear electron path and sample deformation Evaluation of electron micrographs correspondences needs to be done with double precision when using highorder polynomial mappings Non-linear electron projection makes reconstruction computationally intensive. Wide field of view for large datasets CCD cameras are up to 8K x 8K
GPGPUs and CUDA
Guest Lecture, CSE167, Fall 2008
Reconstruction on GPUs
Large datasets take up to several days to reconstruct on a fast serial processor. Goal is to achieve real-time reconstruction Computation is embarrassingly parallel at the tilt level GTX 280 with double-precision support and 240 cores has shown speedups between 10X 50X for large data Tesla units with 4Tflops are the next target for the code.
GPGPUs and CUDA
Guest Lecture, CSE167, Fall 2008
Really ? Free Lunch ?
C-like language support
Missing support for function pointers, recursion, double precision not very accurate, no direct access to I/O Cannot pass structures, unions
Code has to be fairly simple and free of dependencies
Completely self contained in terms of data and variables.
Speedups depend on efficient code
Programmers have to code the parallelism.
No magic spells available for download
Combining CPU and GPU code might be better in cases
GPGPUs and CUDA
Guest Lecture, CSE167, Fall 2008
And more cons
Performance is best for computation intensive apps.
Data intensive apps can be tricky.
Bank conflicts hurt performance Its a black-box with little support for runtime debugging.
GPGPUs and CUDA
Guest Lecture, CSE167, Fall 2008
Resources
http://www.gpgpu.org http://www.nvidia.com/object/cuda_home. html# http://www.nvidia.com/object/cuda_develo p.html http://fastra.ua.ac.be/en/index.html
GPGPUs and CUDA
Guest Lecture, CSE167, Fall 2008
