The programming paradigm provided by CUDA has allowed developers to harness the power of these scalable parallel processors with relative ease, enabling them to achieve speedups of 100 times or more on a variety of sophisticated applications.

The CUDA abstractions, however, are general and provide an excellent programming environment for multicore CPU chips. A prototype source-to-source translation framework developed at the University of Illinois compiles CUDA programs for multicore CPUs by mapping a parallel thread block to loops within a single physical thread. CUDA kernels compiled in this way exhibit excellent performance and scalability. ¹²

Although CUDA was released less than a year ago, it is already the target of massive development activity—there are tens of thousands of CUDA developers. The combination of massive speedups, an intuitive programming environment, and affordable, ubiquitous hardware is rare in today’s market. In short, CUDA represents a democratization of parallel programming. Q

 

REFERENCES

1. NVIDIA. 2007. CUDA Technology; http://www.nvidia. com/CUDA.

2. NVIDIA. 2007. CUDA Programming Guide 1. 1; http:// developer.download.nvidia.com/compute/cuda/1_1/ NVIDIA_CUDA_Programming_Guide_ 1. 1.pdf.

3. Stratton, J.A., Stone, S. S., Hwu, W. W. 2008. M-CUDA: An efficient implementation of CUDA kernels on multicores. IMPACT Technical Report 08-01, University of Illinois at Urbana-Champaign, (February).

4. See reference 3.

5. Buck, I., Foley, T., Horn, D., Sugerman, J., Fatahalian, K., Houston, M., Hanrahan, P. Brook for GPUs: Stream computing on graphics hardware. 2004. Proceedings of SIGGRAPH (August): 777-786; http://doi. acm.org/10.1145/1186562.1015800.

6. Stone, S.S., Yi, H., Hwu, W. W., Haldar, J.P., Sutton, B.P., Liang, Z.-P. 2007. How GPUs can improve the quality of magnetic resonance imaging. The First Workshop on General-Purpose Processing on Graphics Processing Units (October).

7. Stone, J.E., Phillips, J.C., Freddolino, P.L., Hardy, D.J., Trabuco, L.G., Schulten, K. 2007. Accelerating molecular modeling applications with graphics processors. Journal of Computational Chemistry 28( 16): 2618–2640; http://dx.doi.org/10.1002/jcc.20829.

8. Nyland, L., Harris, M., Prins, J. 2007. Fast n-body simulation with CUDA. In GPU Gems 3. H. Nguyen, ed. Addison-Wesley.

9. Golub, G.H., and Van Loan, C.F. 1996. Matrix Computations, 3^rd edition. Johns Hopkins University Press.

10. Buatois, L., Caumon, G., Lévy, B. 2007. Concurrent number cruncher: An efficient sparse linear solver on the GPU. Proceedings of the High-Performance Computation Conference (HPCC), Springer LNCS.

11. Sengupta, S., Harris, M., Zhang, Y., Owens, J.D. 2007. Scan primitives for GPU computing. In Proceedings of Graphics Hardware (August): 97–106.

12. See Reference 3.

Links to the latest version of the CUDA development tools, documentation, code samples, and user discussion forums can be found at: http://www.nvidia.com/CUDA.

 

LOVE IT, HATE IT? LET US KNOW feedback@acmqueue.com or www.acmqueue.com/forums

 

JOHN NICKOLLS is director of architecture at NVIDIA for GPU computing. He was previously with Broadcom, Silicon Spice, Sun Microsystems, and was a cofounder of MasPar Computer. His interests include parallel processing systems, languages, and architectures. He has a B.S. in electrical engineering and computer science from the University of Illinois, and M.S. and Ph.D. degrees in electrical engineering from Stanford University. IAN BUCK works for NVIDIA as the GPU-Compute software manager. He completed his Ph.D. at the Stanford Graphics Lab in 2004. His thesis was titled “Stream Computing on Graphics Hardware,” researching programming models and computing strategies for using graphics hardware as a general-purpose computing platform. His work included developing the Brook software tool chain for abstracting the GPU as a general-purpose streaming coprocessor. MICHAEL GARLAND is a research scientist with NVIDIA Research. Prior to joining NVIDIA, he was an assistant professor in the department of computer science at the University of Illinois at Urbana-Champaign. He received Ph.D. and B.S. degrees from Carnegie Mellon University. His research interests include computer graphics and visualization, geometric algorithms, and parallel algorithms and programming models. KEVIN SKADRON is an associate professor in the department of computer science at the University of Virginia and is currently on sabbatical with NVIDIA Research. He received his Ph.D. from Princeton University and B.S. from Rice University. His research interests include power- and temperature-aware design, and manycore architecture and programming models. He is a senior member of the ACM. © 2008 ACM 1542-7730/08/0300 $5.00