Re:Nvidia and Microsoft team up to build AI cloud computer
by digdugdirk on Sunday November 20, 2022 @12:40AM
(#33678643)
Do Flops still make sense for comparing supercomputers dedicated to machine learning AI tasks? If not, what better ways are there?
[ Parent
| on HN ]
Re:Nvidia and Microsoft team up to build AI cloud computer
by p1esk on Sunday November 20, 2022 @01:26AM
(#33678815)
Not really, though number of flops is still useful to compare workloads (e.g. train resnet50 to a specific accuracy). To compare hardware there are mlperf benchmarks https://mlcommons.org