For more information about this meeting, contact Xiantao Li.
| Title: | How to Design and Program a GPU, and Why You Would Want To |
|---|
| Seminar: | CCMA Luncheon Seminar |
|---|
| Speaker: | Jonathan Cohen, NVIDIA |
|---|
| Abstract: |
|---|
| A modern GPU exploits parallelism to achieve maximum throughput per unit area, rather than maximum single-threaded execution speed. This trend is likely to continue, which will have significant implications for hardware and software designs. NVIDIA's CUDA architecture for GPU Computing provides a fully programmable, massively multithreaded chip with up to 512 scalar processor cores and is capable of delivering performance comparable to supercomputers from only a few years ago. In this talk I will discuss the engineering trade-offs that led to its design. I will trace the evolution from a model single core processor to a massively parallel architecture. I will also cover the basics of the CUDA programming model and how CUDA programs execute efficiently on a GPU. The talk will conclude with recent work that demonstrates the power of GPUs for tackling demanding applications in computational science. |
Room Reservation Information
| Room Number: | MB114 |
|---|
| Date: | 01 / 21 / 2011 |
|---|
| Time: | 12:15pm - 01:30pm |
|---|
