4 ms·
It's highly likely that a workload that is suitable to run on hundreds of disparate computers with thousands of CPU cores is going to be equally well suited for
by wang_li 3y ago
It's highly likely that a workload that is suitable to run on hundreds of disparate computers with thousands of CPU cores is going to be equally well suited for running on tens of thousands of GPU compute threads.
- atq2119 3y agoNot necessarily. GPUs simply aren't optimized around branch-heavy or pointer-chasey code. If that describes the inner loop of your workload, it just doesn't matter how well you can parallelize it at a higher level, CPU cores are going to be better than GPU cores at it.
- monocasa 3y agoThey're not that disparate; the workloads are normally very dependent on the low latency interconnect of most supercomputers.