NVIDIA Vera Rubin MLPerf: What the 2x and 10x Leave Out
Vera Rubin swept MLPerf Training v6.1 collectively, scaled DeepSeek-V3 to 1,024 GPUs at 97% efficiency, and posted a 10x per-watt claim still unverified.

NVIDIA's Vera Rubin posted its first MLPerf numbers, and the headline figures are up to 2x training performance per GPU and 10x agentic inference efficiency per watt. An analysis published on Agentic Affairs argues the fine print is doing most of the work.
The scaling result is the solid one
The cleanest number in the release is about coordination, not peak throughput. DeepSeek-V3 on Vera Rubin scaled from 256 to 1,024 GPUs at roughly 97% efficiency, cutting training time from 16 minutes to about four. The analysis credits NVLink inside the box and Quantum-X800 plus Spectrum-X between boxes, and calls this "a real, concrete result about coordination."
The catch it adds: you bought four times the GPUs to get there, and no pricing or power figures accompany the result, so it is not a claim about cheaper training.
"Up to 2x" and the collective sweep
On the headline training figure, the analysis is blunt: "Up to 2x" is the whole story. Vera Rubin NVL72 delivers up to twice the training performance per GPU versus GB300 NVL72, and "up to" marks the top of a range rather than a uniform doubling. NVIDIA's own account, it notes, does not provide a model-by-model table backing a blanket 2x — so treat it as a ceiling.
The benchmark sweep needs the same reading. NVIDIA swept every MLPerf Training v6.1 benchmark collectively, across Vera Rubin, GB300 and GB200 submissions. That is a statement about bench depth, not proof that Vera Rubin alone won every race — a distinction that matters if you are the one buying hardware.
The 10x agentic claim is provisional
The efficiency headline rests on a Kimi K3 agentic inference comparison, and the analysis lists four caveats, noting NVIDIA itself marks the result as pending MLCommons verification.
| Measure | Vera Rubin | GB300 |
|---|---|---|
| Throughput | 41,414 tokens/sec | 2,210 tokens/sec |
| Concurrency | 1,024 queries | 48 queries |
| Accelerators | 32 | 48 |
| P90 time to first token | 4,674 ms | 2,305 ms |
| Power at measured point | 100.89 kW | 60 kW |
Two of those rows cut against the marketing. Vera Rubin's P90 time to first token is worse, so "same response speed" turns out to mean both configurations cleared a shared minimum interactivity floor. And the Vera Rubin configuration drew more power at the measured point — the win is throughput relative to power, not a smaller electricity bill.
What the benchmark does not say
MLPerf Training measures time to a target accuracy. Nothing in these results shows an agent became more accurate, made better decisions, or finished more useful work. Faster training shortens one slice of development; whether it shortens your release cycle depends on everything the benchmark does not measure.
The analysis lands on a split verdict worth carrying into a procurement conversation: "Buy the scaling story. Rent the 10x story until MLCommons signs off." If you are sizing a cluster on these numbers, the 97% scaling efficiency is the figure to plan against, and the per-watt agentic claim is the one to revisit after verification — along with the pricing and power data the debut announcement does not include.
More from DangMua