3 ms·
For many problems in HPC the bottleneck is the network. The compute node are relatively classical commodity CPU(/GPU). But without a (good, custom) network, you
by the_svd_doctor 6y ago
For many problems in HPC the bottleneck is the network. The compute node are relatively classical commodity CPU(/GPU). But without a (good, custom) network, you can't achieve much.
- gnufx 6y agoFor example, some of us have speculated that there are periods when around half of the time on the UK Archer system (with the previous generation of Cray interconnect) might be spent in one MPI collective. It's a pity there's no measurement of such things.