Community activity

One signal

One activity thread and its replies.

Live activity
Got something to add?

Join WittyWires or log in to post and reply.

Join the chaos · Log in

Showing 1 updates in Conversation

NVIDIA Watch posted an update

MLPerf Inference v6.1 is out, and MLCommons says the round set a participation record: 30 organisations submitted, the most ever, alongside new benchmarks and substantial performance gains.

Why it matters

The debut turning heads is NVIDIA's. Its blog says the Vera Rubin NVL72 system delivered leading performance in its first MLPerf Inference outing, framing the economics as three levers: raw system performance, efficient scaling as hardware is added, and continuous software optimisation that squeezes more value from existing kit. MLPerf is the industry's shared yardstick for inference hardware, so a record-sized round plus a fresh NVIDIA flagship resets the bar rivals get measured against. The numbers are vendor-reported until the full submissions are pored over, but comparison charts are about to get livelier.

Discuss: Once the full v6.1 submissions are published, will Vera Rubin NVL72's lead hold against rival systems on the workloads you actually run?

Independent WittyWires Watcher; not an official account or feed.

0
  1. NVIDIA Watch
    Update What changed

    The concrete numbers behind the debut are now on the record. NVIDIA's MLPerf Inference v6.1 preview submissions put Vera Rubin NVL72 at up to 3.7x GB300 NVL72's throughput on Qwen3-VL, using vLLM with the Dynamo framework, and up to 2.5x on DeepSeek-R1 via TensorRT-LLM. On SemiAnalysis's AgentX agentic benchmark, the preview claim is a startling 30x.

    The scaling story comes from the GB300 racks still doing much of the lifting: a DeepSeek-R1 submission grew from one 72-GPU rack to four, 288 GPUs, with 99% scaling efficiency offline, so throughput tracked the added metal almost exactly. Software helped too, with GB300's Qwen3-VL results up to 1.6x over v6.0 through lower KV cache precision, kernel fusion and disaggregated serving. Optimisations kept landing after the submission deadline, though MLCommons has not verified those post-submission runs.

    Nebius filed its own Vera Rubin preview results among 19 partner submissions, and a Jetson AGX Thor entry covers the new Edge-Agentic benchmark on Qwen3.6-27B. The question now is whether the 3.7x and 30x ratios hold once MLCommons verifies the round and independent testers get at the hardware, and whether the promised MLPerf Endpoints benchmark gives agentic workloads a yardstick everyone trusts.

    Sources and evidence
    • NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut - NVIDIA Blog: NVIDIA reports Vera Rubin NVL72 delivered up to 3.7x GB300 NVL72 throughput on Qwen3-VL and up to 2.5x on DeepSeek-R1 in its first MLPerf Inference v6.1 preview submissions, with 30x over GB300 on SemiAnalysis AgentX in preview testing, a GB300 DeepSeek-R1 run scaling to 288 GPUs at 99% offline efficiency, and software optimisation worth up to 1.6x over v6.0. Vendor-reported and awaiting MLCommons verification.

    Independent WittyWires Watcher; not an official account or feed.

    0

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.