Blackwell Sets MLPerf Records
NVIDIA’s Blackwell Ultra platform posted record-setting MLPerf Inference v6.0 results across every category and scenario, becoming the first submitter for all new model/test cases. The sweep underscores NVIDIA’s lead in inference for transformer and recommendation workloads. (wccftech.com) (storagereview.com)
MLPerf Inference v6.0 adds five new or updated datacenter tests — including an open-weight GPT-OSS-120B LLM, an expanded DeepSeek‑R1 interactive benchmark, the transformer‑based recommender DLRMv3, a text‑to‑video WAN‑2.2 test, and a new Shopify-derived vision‑language model — as part of the suite’s April 1, 2026 release. (mlcommons.org) NVIDIA’s submission used scale‑out inference with four GB300 NVL72 rack systems interconnected by Quantum‑X800 InfiniBand, totaling 288 Blackwell Ultra GPUs, to exercise system‑level throughput at MLPerf v6.0. (developer.nvidia.com) That 288‑GPU setup posted 2,494,310 tokens/sec in the DeepSeek‑R1 offline scenario and 1,555,110 tokens/sec in server mode, according to the published MLPerf v6.0 results table. (storagereview.com) Published MLPerf figures for other v6.0 models show GPT‑OSS‑120B at 1,046,150 tokens/sec offline and 1,096,770 tokens/sec server, Qwen3‑VL at 79 samples/sec offline, and DLRMv3 at 104,637 samples/sec offline and 99,997 queries/sec in server mode. (storagereview.com) NVIDIA attributes the performance uplift to full‑stack optimizations — TensorRT‑LLM and Dynamo improvements plus techniques such as kernel fusion, optimized attention data parallelism, disaggregated serving, Wide Expert Parallel, Multi‑Token Prediction and KV‑aware routing — claiming up to 2.7× throughput gains and over 60% lower cost per token on the same infrastructure. (developer.nvidia.com) The company reports cumulative MLPerf training and inference wins since 2018 of 291 (about nine times the total of other submitters) and says 14 partners including ASUS, Cisco, CoreWeave, Dell, Google Cloud, HPE, Lenovo and Supermicro submitted results on NVIDIA‑based systems in this round. (developer.nvidia.com)
