AI
•3 min read
Meta Details Prometheus Gigawatt AI Supercluster Scaling Toward 1M GPUs in Ohio
Zubiqo Take
“Expanding the NVLink domain from 8 to 72 GPUs multiplied the hardware fault blast radius by 9x, meaning Meta has to leave roughly 13% of its provisioned Blackwell GPUs sitting in reserve just to keep a single bad component from halting an entire training run.”
AI: BullishMag 8
Engineering at Meta