NVIDIA says Blackwell platform led every category in MLPerf Training 6.0 benchmarks

NVIDIA is claiming a clean sweep in the newest MLPerf Training 6.0 benchmark round, a peer-reviewed industry test designed to measure how quickly and reliably AI training jobs can be run at scale. In a company blog post published June 16, NVIDIA said it led across every category, attributing the results to what it calls extreme co-design between its Blackwell hardware, networking, and training software stack, aimed at letting AI builders train frontier models faster while keeping runs repeatable.

MLPerf Training 6.0, the latest version of the benchmark suite, added two new mixture-of-experts (MoE) pretraining workloads this round: DeepSeek-V3 at 671 billion parameters and GPT-OSS at 20 billion parameters. Mixture-of-experts models route different parts of a model to specialized “expert” subnetworks, which typically increases model capacity without requiring every expert to be active for every token, but it also raises heavy communication and scheduling demands across GPUs.

NVIDIA said it was the only vendor submitted across every benchmark in the suite, and that it delivered the fastest time to train on all seven workloads. For the benchmark configuration, NVIDIA reported results from its rack-scale Blackwell systems, including NVIDIA GB200 NVL72 and GB300 NVL72. NVIDIA’s rack-scale setup uses fifth-generation NVIDIA NVLink Switches to connect 72 GPUs into a unified pool of compute and memory, allowing them to operate as a single large training system.

NVIDIA also highlighted performance differences between the two generations. It said GB300 NVL72 delivered up to 1.6 times faster training than GB200 NVL72 at the same scale in this round. The company linked the gain to a set of Blackwell Ultra capabilities, including higher compute density enabled by NVFP4 training methods, expanded memory capacity, and what it described as a higher power ceiling that helps the system sustain peak performance during long-running training.

On the MoE workloads specifically, NVIDIA pointed to the communication bottleneck created by expert routing, saying that training MoE models involves all-to-all communication similar to MoE inference, because tokens need to be directed across GPUs to reach the correct expert subnetwork. NVIDIA attributed its scaling speed at large cluster sizes to NVLink’s bandwidth advantage, framing it as critical for keeping expert routing fast and efficient as the number of GPUs grows.

Beyond raw throughput, NVIDIA said the MLPerf results also reflect resiliency, or the ability to produce reproducible training outcomes over time. The company described its platform as being engineered along two “dimensions” to support reliability at training scales where runs can take weeks or months across hundreds of thousands of GPUs, though it did not provide additional specific metrics or failure-rate data in the blog post.

NVIDIA’s submissions included scaling MoE training to large GPU counts. For DeepSeek-V3 671B, NVIDIA said it scaled to 8,192 GPUs using GB200 NVL72 systems, describing that as the largest-scale Blackwell-based submission in MLPerf Training to date. It also reported a result for Llama 3.1 405B, a dense large language model, at 5,120 GPUs using GB200 NVL72 systems, characterizing the run as one of the largest dense LLM entries in the suite.

Alongside the GPU systems, NVIDIA emphasized complementary networking options for distributed training, naming NVIDIA Quantum InfiniBand and NVIDIA Spectrum-X Ethernet as “scale-out” platforms data centers can use to build large clusters tuned to their infrastructure. The company also cited partner co-engineering across system architecture, networking, and software, listing 19 organizations that participated in the round with submissions, including cloud providers, hardware vendors, and AI infrastructure firms. In addition to those partner submissions, NVIDIA highlighted several customer or partner efforts, including Cohere’s claim of 3 times faster training on GB200 NVL72 for its North agentic AI platform, Midjourney’s scaling of Blackwell Ultra GPU fleets on CoreWeave, and Google Cloud’s Thinking Machines Lab reporting 2 times faster training and serving speeds on GB300 NVL72 compared with prior-generation GPUs. NVIDIA also referenced Nebius, saying it reported a 30% reduction in model training time for Higgsfield and described a downstream output scale for its platform.

Still, some details remain unclear from the blog post alone. NVIDIA did not publish the benchmark score breakdowns, per-workload configurations, or training accuracy verification values in the excerpted text, and the company did not specify exactly which system-level parameters (batch sizes, precision settings, or run-to-run variance) produced the “reliability” outcomes it referenced. Readers looking for full verification would need to consult the MLPerf Training 6.0 published results pages and technical notes, which would typically include the exact run parameters and rule compliance details.

Going forward, the market will likely watch two things. First, whether NVIDIA’s reported performance edge translates into measurable reductions in time-to-results for customers training frontier models, especially for MoE workloads where communication overhead can dominate. Second, the competitive response from other GPU and system vendors, particularly as MLPerf Training 6.0 expands MoE coverage and increasingly tests the end-to-end stack, not just compute throughput.