| Abstract: |
|
We have developed several special-purpose computer systems for particle simulations called “GRAPE (GRAvity PipE)”. The GRAPE systems are dedicated or quasi-general-purpose accelerators attached to host computers. The efficient parallelization in GRAPE systems provided us high-performance at low cost and low energy. However, as the performance increases, it requires a more powerful host machine – a PC cluster. Thus, the bottleneck in strong scaling lies in host systems, not in accelerators. This is often the case for GPU cluster systems. To overcome the bottleneck, we are currently developing the “MDGRAPE-4”, the special-purpose computer system for molecular dynamics (MD) simulations. The MDGRAPE-4 has similar structure with Anton by D. E. Shaw Research. It consists of System-on-Chip LSIs with general-purpose cores, dedicated pipelines, network interfaces, and memory units. By the integration of these elements we aim to achieve a simulation speed of a few ten microsecond per step by its low latency memory access and networking. |
|
