Build1 publisher2 min readPublished
NVIDIA's 3.7x Vera Rubin figure comes from Qwen3-VL on vLLM and Dynamo
MLPerf Inference v6.1 preview submissions put Vera Rubin NVL72 at up to 3.7x GB300 on Qwen3-VL and up to 2.5x on DeepSeek-R1, on two different inference frameworks. The four-rack 99% scaling result is an offline number.
The Engineer · Build desk
What happened
- NVIDIA's MLPerf Inference v6.1 post reports Vera Rubin NVL72 preview submissions on two of the suite's hardest benchmarks, DeepSeek-R1 and Qwen3-VL, the system's first MLPerf appearance.
- The headline claim of up to 3.7x GB300 NVL72 throughput is a Qwen3-VL result spanning the offline, server and interactive scenarios, run on vLLM with NVIDIA's Dynamo framework.
- On DeepSeek-R1, run with the TensorRT-LLM library instead, NVIDIA reports up to 2.5x the throughput of GB300 NVL72.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision The 3.7x came from vLLM plus Dynamo, the 2.5x from TensorRT-LLM. A buyer comparing racks has to match the claim to the stack it was measured on, and a shop standardised on one framework cannot carry the other's multiplier into its sizing model.
- constraint The 99% efficiency result covers the offline scenario, so anyone planning four racks for latency-bound serving still has to measure server and interactive scaling on their own traffic before committing.
- cost Software moved throughput up to 1.6x inside a single benchmark round. Throughput written into a purchase or a capacity plan ages faster than the hardware it describes.
- precedent Until MLPerf Endpoints exists, agentic performance claims such as the 30x AgentX result sit on a benchmark a buyer cannot line up against other vendors' submissions.
Take the 3.7x apart and it is one model on one software stack. NVIDIA got it on Qwen3-VL, running vLLM with its Dynamo inference framework [2]. The DeepSeek-R1 figure, up to 2.5x, came out of TensorRT-LLM [3]. Same rack, two frameworks, and the larger multiplier is 48% above the smaller one [15].
The techniques are named in the post. NVFP4 precision cuts the memory footprint of model weights, attention and the KV cache, and NVIDIA says the throughput gain arrives with "minimal loss of output quality" [7]. The submissions used disaggregated serving, splitting prefill from decode, plus large-scale expert parallelism across the mixture-of-experts layers in DeepSeek-R1 and Qwen3-VL [8]. Both of those techniques lean hard on the fabric, and NVIDIA credits sixth-generation NVLink and NVLink Switch with 10x higher packet rates and 3x lower latency than off-the-shelf Ethernet [9].
For the 3.7x to transfer, your model has to be a mixture of experts wide enough to reward expert parallelism, and your serving stack has to separate prefill from decode [8]. NVFP4 has to hold quality on your own evals, and NVIDIA quantifies the loss only as "minimal" [7]. The topology has to be the full 72-GPU NVLink scale-up domain [9].
The scaling submission is easier to read. NVIDIA scaled DeepSeek-R1 from one GB300 NVL72 rack at 72 GPUs to four racks at 288 GPUs and reported 99% scaling efficiency in the offline scenario [5]. At 99%, four racks return 3.96 racks of throughput, so roughly 2.9 GPUs' worth of the 288 goes to overhead [14]. The 3.7x, by contrast, is quoted as an up-to spanning offline, server and interactive together [2].
Every figure here comes from NVIDIA's own post describing NVIDIA's own submissions [17]. The other Vera Rubin NVL72 preview submitter the post names is Nebius, introduced as part of the partner ecosystem [10]. The 30x agentic result is from SemiAnalysis AgentX in preview testing, not from the MLPerf suite [11]. MLPerf's own agentic benchmark, Endpoints, is described as upcoming [12].
The software line is the one I would price into a purchase decision. NVIDIA says optimizations lifted its v6.1 submissions up to 1.6x over v6.0, and that further gains landed after the v6.1 submission closed [6]. Anyone who signed against v6.0 throughput got part of that without buying hardware.
NVIDIA also writes that each Vera Rubin NVL72 rack "delivers significantly more tokens, serves more users and generates more revenue than a GB300 NVL72 rack, while lowering cost per token" [16]. The post does not give rack power or price. Without those two inputs, a buyer cannot recompute cost per token from the published results.
What to watch
- Whether the published v6.1 tables break Vera Rubin out per scenario for server and interactive, instead of a single up-to maximum.
- Whether MLPerf Endpoints, once it exists, reproduces anything near the 30x AgentX gap on agentic workloads.
- Whether a submitter outside NVIDIA and Nebius posts Vera Rubin NVL72 results with rack power attached.