Skip to content
trendoviz

NVIDIA says Vera Rubin leads MLPerf benchmarks

NVIDIA reports that Vera Rubin led its submitted MLPerf Training and Endpoints results in its benchmark debut. The announcement includes an all-benchmark training sweep and generation-over-generation efficiency comparisons against earlier NVIDIA systems, but no independent result tables are provided.

47 out of 100Emerging7D: Insufficient history ProvisionalPublished

Overview

NVIDIA’s announcement says Vera Rubin led performance across training, post-training and agentic inference in its MLPerf debut [1]. According to NVIDIA, the platform won every benchmark in MLPerf Training v6.1 and recorded more than 10 times the previous generation’s throughput per watt in its first MLPerf Endpoints submission [1].

These are vendor claims, not independently verified results. The announcement does not provide the underlying benchmark tables or competing submissions needed to check the claimed sweep or compare every system [1].

The generation-over-generation figures are narrower. NVIDIA reports up to twice the performance per GPU for Vera Rubin NVL72 compared with GB300 NVL72 in its debut training results. In a provisional Kimi K3 endpoint result, it reports more than 10 times GB300 NVL72’s throughput per watt at the same response speed [1]. Neither comparison establishes performance against non-NVIDIA platforms [1].

Why it matters

The reported training figures point toward shorter model-development cycles. If the scaling behavior holds under independently checked configurations, teams that repeatedly retrain or post-train models could spend less time waiting for each run [1].

The endpoint results address a different constraint: serving more agent activity within a fixed power envelope. NVIDIA describes throughput per watt as the amount of agentic work completed per watt, while the endpoint benchmark also checks task completion [1]. These claims may matter to operators where power, rather than access to accelerators, limits deployment density [1].

Opportunities

  • Evaluate training iteration needs. Teams working on large models can assess whether shorter training cycles would materially improve their model-update workflow, using Rubin as a reference point rather than an assumed default [1].
  • Test agent-serving density. Power-constrained operators can include throughput per watt and task completion in their own evaluations of agent-serving configurations [1].
  • Request comparable evidence. Buyers can ask for matched MLPerf Training and Endpoints configurations, including absolute throughput, power consumption, accelerator counts and software versions [1].

Risks

NVIDIA’s comparisons cover earlier NVIDIA configurations, including GB300 NVL72 and Azure’s GB200 systems. They do not establish how Rubin compares with non-NVIDIA platforms [1].

The Kimi K3 endpoint comparison is explicitly provisional and may change [1]. NVIDIA also expects future gains from continued optimization of its CUDA software stack, which means reported performance can depend on configuration and software maturity [1].

A heavy reliance on NVIDIA-specific libraries, interconnects and tooling could make migration more difficult. The announcement does not quantify those dependencies or their costs [1].

Context

MLPerf Training measures how quickly a system trains models to specified accuracy targets across workloads such as language models, image generation, recommender systems and reinforcement-learning post-training [1]. MLPerf Endpoints evaluates agent-serving systems on coding and enterprise workflows and checks that agents complete their tasks correctly [1].

NVIDIA’s DeepSeek-V3 example concerns training scale rather than the claimed benchmark sweep. The company says Vera Rubin NVL72 scaled from 256 to 1,024 GPUs with about 97% efficiency, reducing training time from 16 minutes to about four minutes [1]. NVIDIA also reports running the benchmark at 10,240 GPUs with GB300 systems and at the same scale on Microsoft Azure using GB200 systems [1].

NVIDIA says 20 organizations submitted MLPerf Training results, including ASUS, Azure, Cisco, CoreWeave, Crusoe, Dell Technologies, Firmus, Giga Computing, Google Cloud, HPE, Inventec, Krai, Lambda, Nebius, Oracle, QCT, Supermicro, TTA and the University of Florida [1].

Known, claimed, inferred

The evidence establishes that NVIDIA submitted Vera Rubin to MLPerf Training v6.1 and MLPerf Endpoints and that 20 partner organizations also submitted training results [1]. The central leadership, sweep and efficiency claims come from NVIDIA’s own announcement rather than a third-party test report [1].

The DeepSeek-V3 result is the clearest training-scale example in the announcement: NVIDIA reports that quadrupling the GPU count from 256 to 1,024 cut training time from roughly 16 minutes to four at about 97% scaling efficiency [1]. The announcement does not include logs or independent validation for that figure [1].

The endpoint comparison adds a power-efficiency dimension, but the headline multiple is tied to a provisional result and to NVIDIA’s previous generation. The announcement also says Rubin led SemiAnalysis AgentX and Artificial Analysis AgentPerf, but it supplies no underlying results from either benchmark, so those claims are not independently verified here [1].

Limitations & uncertainties

There is one substantive source: NVIDIA’s announcement. A second NVIDIA post is a truncated repetition and adds no corroboration [1].

Readers evaluating the claims should request benchmark-level scores, system configurations, accelerator counts, software versions, power measurements and task-completion criteria from vendors or MLCommons. They should also ask whether result tables cover all participating systems and how provisional results changed after finalization [1].

The announcement does not define the methodology behind “leading performance,” “swept every benchmark,” “up to 2x” or “more than 10x.” Absolute throughput, power and cost figures for competing systems are absent [1].

Evidence & sources

The key statements of this analysis, labeled by type, with the sources they rely on. Trendoviz does not run product tests.

  • Primary-source or vendor statementNVIDIA says Vera Rubin led performance across training, post-training and agentic inference in its MLPerf debut.[1]
  • Primary-source or vendor statementNVIDIA says Vera Rubin won every benchmark in MLPerf Training v6.1.[1]
  • Primary-source or vendor statementNVIDIA reports up to twice the performance per GPU for Vera Rubin NVL72 compared with GB300 NVL72 in its debut training results.[1]
  • Primary-source or vendor statementNVIDIA reports that Vera Rubin NVL72 scaled from 256 to 1,024 GPUs at about 97% efficiency, cutting DeepSeek-V3 training time from 16 minutes to about four.[1]
  • Primary-source or vendor statementNVIDIA says its first MLPerf Endpoints submission delivered more than 10 times the previous generation’s throughput per watt for agentic AI.[1]
  • Primary-source or vendor statementNVIDIA reports more than 10 times GB300 NVL72’s throughput per watt at the same response speed in a provisional Kimi K3 endpoint result.[1]
  • Primary-source or vendor statementNVIDIA says 20 organizations submitted MLPerf Training results.[1]
  • Primary-source or vendor statementNVIDIA says Vera Rubin led results on SemiAnalysis AgentX and Artificial Analysis AgentPerf, but the underlying data are not present.[1]

Sources

  1. 1

Updates & corrections

  1. First published · published automatically after passing every check

Found an error? Report it.