2 ms·
Well I guess the simulation would have to be broken up to run on multiple GPUs at once anyway. But even if you had a good interconnect, the GPUs in this box (K4
by sp332 12y ago
Well I guess the simulation would have to be broken up to run on multiple GPUs at once anyway. But even if you had a good interconnect, the GPUs in this box (K40) have almost 2x as many processors and 3x as much RAM. (Edit: compared to AWS GPU instances.) If you're doing a 3D sim, each node needs to communicate with 26 others, so just having 1/3 as many nodes is going to help a lot.