GPU acceleration
GPU accelerated simulations
27.04.2012
- Simulations of magnetic reconnection with 8192 x 4096 cells, 100 particles per cell over 600000 time steps have been performed in less than 6 hours wall clock time on 512 GPUs on Jaguar at ORNL. Jaguar presently has 900 nodes with GPU accelerators. Jaguar will be upgraded to a few thousand GPUs in the near future.
- 2D simulations on up to 900 GPUs have shown good performance scaling. About 215 million particles per second and GPU, including all sorting, boundary exchange, etc. have been processed. On Jaguar one GPU is coupled with one 16-core CPU. There is a lot of CPU computing power too. In fact, up to about 160 million particles per second and CPU can be processed on the CPU.