๐ November 16, 2026 โข NVIDIA
Vera Rubin debuts with a clean sweep in MLPerf
NVIDIA published the results on November 16, 2026, alongside SC26, the annual supercomputing conference held this week in Chicago (per NVIDIA). The company reports that Vera Rubin swept every MLPerf Training v6.1 benchmark in its debut, the only platform with a result in each test of the round, which covered LLM workloads, image generation, recommender systems and RL post-training (per NVIDIA). MLPerf is run by MLCommons as an independent consortium, and results arrive as vendor submissions under its rules (per MLCommons).
Up to 2x per GPU and scaling to 10,240 GPUs
Vera Rubin NVL72 delivered up to 2x more performance per GPU than an optimized GB300 NVL72, early generation-over-generation results (per NVIDIA). On a large open LLM benchmark it scaled from 256 to 1,024 GPUs with about 97 percent efficiency, cutting training time from 16 to about 4 minutes (per NVIDIA). The largest submission of the round hit 10,240 GPUs on GB300, matched by Microsoft Azure with 10,240 GPUs on GB200 (per NVIDIA). NVIDIA credits NVLink scale-up, Quantum-X800 InfiniBand and Spectrum-X Ethernet scale-out, with software keeping thousands of GPUs aligned (per NVIDIA).
First MLPerf Endpoints run measures agentic work
MLPerf Endpoints tests systems on real agentic coding and enterprise workflows and verifies that agents complete tasks correctly (per NVIDIA). In a provisional result on a large open agentic model, NVIDIA reports that Vera Rubin systems deliver more than 10x throughput per watt versus GB300 NVL72 at the same response speed (per NVIDIA). The submission used 32 accelerators serving 1,024 parallel requests, with 41,414.2 tokens per second of system throughput, 50.4 tokens per second per user and 100.89 kW of power draw (per NVIDIA). Endpoints v1.0 appeared on November 9, 2026 (per NVIDIA).
Rubin HPC brings a record CUDA core count to science
NVIDIA also announced Rubin HPC at SC26, a new GPU in the Vera Rubin platform for simulations and data processing in science, engineering and analytics (per NVIDIA). It carries more than 50,000 CUDA cores on one GPU, a record, with sparse data acceleration and GPU-to-GPU links over seventh-generation NVLink (per NVIDIA). The design balances data movement with compute, speeding sparse matrix operations and stencil computations (per NVIDIA). DataCenterDynamics and TomHardware reported the same record core count, and NVIDIA expects availability in 2028 (per NVIDIA).
Faster discovery, and how to read the numbers
CUDA-X agent libraries, cuPhoton, cuPotential and cuSpa, speed the loop of hypothesis, simulation and refinement (per NVIDIA). Ian Buck, vice president of hyperscale and HPC at NVIDIA, tied faster simulation to faster discovery for agents in science and engineering (per NVIDIA). The benchmark figures are vendor submissions under MLCommons rules on NVIDIA and Azure systems with NVIDIA GPUs, which NVIDIA calls early results, with software optimization set to add more (per NVIDIA). We track them as vendor statements within MLCommons methodology, not an independent verdict (per MLCommons).
Sources:
blogs.nvidia.com/blog/mlperf-training-endpoints-vera-rubin
blogs.nvidia.com/blog/rubin-hpc
datacenterdynamics.com
tomshardware.com
mlcommons.org
Continue the conversation: https://t.me/AIAutomationArchitecture ยท Bluesky
Top comments (0)