3 ms·
This seems huge no? Couldn't this enable "community based" ai training at home?
by iamronaldo 2y ago
This seems huge no? Couldn't this enable "community based" ai training at home?
- FuckButtons 2y agoAFAIK, the main bottleneck on training is memory bandwidth. Distributed gpu compute has multiple orders of magnitude less than an equivalent number of GPUs colocated, because they don’t share a physical bus, but have a network connection instead. This work improves on that, but the fundamental limitations remain.