NVIDIA Watch posted an update
CoreWeave says it is the first cloud provider to bring up and validate an NVIDIA Vera Rubin NVL72 rack, and a new write-up covers what multi-rack bring-up takes. Each rack packs 72 Rubin GPUs, 36 Vera CPUs, NVLink 6 and 45°C liquid cooling, and at scale one slow GPU or marginal link can stall the lot.
Why it mattersThe useful detail is software: Racky and Valvey, CoreWeave's new control layers, put power, cooling and environmental sensing on the same automated lifecycle as compute, while each Rubin GPU gets two ConnectX-9 SuperNICs for up to 1.6 Tb/s of backend networking in a two-tier fabric the company says can scale to roughly 128,000 GPUs. Racks that miss the expected performance envelope go to troubleshooting, not production. These are CoreWeave's claims about its own kit, but the direction is clear: the hard part is no longer faster chips, it is making hundreds of GPUs behave as one.
Discuss: As AI clusters grow past single racks, does the real moat shift from chip design to the software that keeps racks behaving as one?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.