3 ms·
Supercomputing environments I've encountered are still timesharing-based, because by definition they are too expensive to be single-user. My old university stil
by kedean 3y ago
Supercomputing environments I've encountered are still timesharing-based, because by definition they are too expensive to be single-user. My old university still operates a cluster-based system that you can submit jobs to, with "head nodes" that have ssh access for light work/setup (they aren't suited for heavy jobs). It's operated by the CS dept but available for others to use, the sciences in particular.
- sneed_chucker 3y agoFrom what I've seen at FAANG companies, GPU clusters for ML researchers are similar. Basically a bunch of beefy nodes with 96 CPUs and 8 Nvidia A100 cards each configured in a SLURM cluster. Users get SSH access and can run tmux, jupyter, or whatever else is needed for their workflow.