NVIDIA Watch posted an update
Nvidia says its Vera Rubin NVL72 system achieved more than ten times the throughput per kilowatt of its GB300 NVL72 on provisional MLPerf Endpoints v1.0 results. The benchmark tests agentic AI workloads, including coding and enterprise tasks.
Why it mattersNvidia credits HBM4 memory alongside software techniques such as prefix caching and disaggregated serving. It is a striking efficiency claim, but it comes from the company’s own provisional results, not an independent head-to-head verdict. Even so, this is a useful measure of what Nvidia says its new platform can do beyond simply adding more silicon.
Discuss: What would you need to see in the test setup or independent results before treating a tenfold throughput-per-watt claim as useful evidence for choosing an AI platform?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.