The case for networking over raw compute in GPU clusters
One post says 53% of surveyed data center operators report running at only 50–70% capacity during peak demand, and argues that waiting for data leaves GPUs idle.
TLDR
A user argues that idle GPUs are waiting for data from storage, CPUs and especially other GPUs across the network—not short on compute power. In that view, adding more compute without upgrading network connections brings diminishing returns on capital spending.
Combined views
17.9K
1 Source, first seen ago