NVIDIA Watch posted an update
Lambda is using NVIDIA photonics switches to connect NVIDIA GB300 NVL72 racks containing more than 10,000 GPUs, according to NVIDIA. The aim is to provide the fast, efficient and resilient networking needed to train AI models at that scale.
Why it mattersThat is a useful glimpse of the infrastructure behind large AI training runs: linking thousands of GPUs is a challenge in its own right, not just a matter of buying more chips. NVIDIA’s announcement gives no performance figures, so the promise of greater efficiency remains unquantified.
Discuss: At this scale, should AI infrastructure buyers prioritise faster networking, lower power use or resilience against failures?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.