Supercomputing & AI Infrastructure
NVIDIA Unveils Rubin HPC GPU and Sweeps MLPerf as Lambda Deploys CPO
NVIDIA introduced the 50,000-core Rubin HPC GPU at SC26, swept every MLPerf Training v6.1 benchmark with over 10x agentic throughput per watt on Vera Rubin NVL72, and supplied co-packaged Photonics switches for Lambda's 10,000-GPU cluster.

Your listening Queue & catch up
Add a story or build a catch-up queue to start listening.
High-performance scientific simulation, large-scale model training, and data center optical networking advanced on Monday at the SC26 supercomputing conference as NVIDIA unveiled its Rubin HPC GPU architecture, published sweeping benchmark results across MLPerf Training v6.1 and MLPerf Endpoints v1.0, and detailed cloud provider Lambda's deployment of co-packaged optical switches across more than 10,000 GPUs.
NVIDIA Unveils 50,000-Core Rubin HPC GPU for Scientific Simulation
Announced Monday at SC26, the NVIDIA Rubin HPC is a new GPU in the NVIDIA Vera Rubin platform engineered to accelerate simulation and data processing across science, engineering, and analytics workflows, according to the NVIDIA Blog. The Rubin HPC GPU features a record count of more than 50,000 CUDA cores on a single GPU, paired with sparse data acceleration to keep those cores supplied with data and seventh-generation NVIDIA NVLink interconnect technology for GPU-to-GPU connectivity. Expected to be available in 2028, the processor is designed for researchers using AI agents to explore candidate materials, molecules, and industrial designs before committing to physical prototypes.
Ian Buck, vice president of hyperscale and high-performance computing at NVIDIA, stated that faster simulation translates directly into faster discovery for agents tackling science and engineering problems. Across climate models, engineering simulations, genomic analysis, and financial risk calculations, Rubin HPC scales data movement and compute together to accelerate sparse matrix operations, stencil computations, sorting, and reductions. Working alongside NVIDIA CUDA-X libraries, the architecture shortens the feedback loop of hypothesis, simulation, and refinement for laboratory teams.
Vera Rubin Platform Sweeps MLPerf Training v6.1 and Endpoints v1.0
In its benchmark debut released Monday, the NVIDIA Vera Rubin platform swept every MLPerf Training v6.1 test and led the inaugural MLPerf Endpoints v1.0 evaluations for agentic inference, the NVIDIA Blog reported. Spanning large language models, image generation, recommender systems, and reinforcement learning post-training, Vera Rubin NVL72 delivered up to 2x more performance per GPU than the NVIDIA GB300 NVL72 system. On the DeepSeek-V3 training benchmark, Vera Rubin NVL72 scaled from 256 to 1,024 GPUs with about 97% efficiency, cutting training time from 16 minutes to about 4 minutes. NVIDIA also executed the benchmark across 10,240 GPUs on GB300 systems—the largest scale in the round—which Microsoft Azure matched using 10,240 NVIDIA GPUs on GB200 systems linked by NVLink, Quantum-X800 InfiniBand, and Spectrum-X Ethernet networking.
In MLPerf Endpoints v1.0 testing on the open Kimi K3 model for agentic workflows, a 32-accelerator Vera Rubin NVL72 system delivered more than 10x the throughput per watt of a 48-accelerator NVIDIA GB300 NVL72 system at a shared minimum interactivity threshold of 50 tokens per second per user. Handling a concurrency of 1,024 queries at 100.89 kW operating power, Vera Rubin NVL72 achieved 41,414.2 tokens per second and 50.4 tokens per second per user with a 90th-percentile time to first token of 4,673.9 ms, compared with GB300 NVL72's 2,210.4 tokens per second across 48 concurrent queries at 60 kW and 2,305.1 ms. Submissions in the round included 20 partner organizations, among them ASUS, Azure, Cisco, CoreWeave, Crusoe, Dell Technologies, Google Cloud, HPE, Lambda, Nebius, Oracle, and Supermicro.
Lambda Connects Over 10,000 GPUs With NVIDIA Photonics Switches
To eliminate networking bottlenecks in large training and inference clusters, cloud provider Lambda is deploying NVIDIA Photonics co-packaged optics switches to connect NVIDIA GB300 NVL72 racks housing more than 10,000 GPUs, according to an NVIDIA Blog announcement. Dave Ward, president of cloud services at Lambda, noted that integrating co-packaged optics at this scale supports next-generation AI infrastructure by lowering power draw and extending mean time between interruptions. In Lambda's testing, network latency dropped by about 4% at rack scale and by up to 10% at pod scale while maintaining full throughput across rack, scalable unit, and pod configurations.
By placing optical engines directly on the switch package and shortening electrical paths from inches to millimeters without retimer chips, NVIDIA Photonics delivers up to 5x better power efficiency and up to 15x higher reliability than conventional pluggable optics while using 4x fewer lasers. Each liquid-cooled switch features 144 ports operating at 800Gb/s and 18 field-replaceable external light sources, holding median optics temperatures to 36°C compared with 51°C for conventional pluggable transceivers.
Sources
Follow the coverage
Published reporting connected by NVIDIA.
- CrowdStrike Acquires Fiddler and Dreadnode as SOCs and Cities Scale AI
- NVIDIA Unveils Rubin HPC GPU and Sweeps MLPerf as Lambda Deploys CPOYou are reading this story
- NVIDIA Details Supply Chain as Apple Tests Siri Caps & OpenAI Grows
- UC Berkeley Joins Schmidt AI Network as NVIDIA Launches DGX Spark Kit
Comments (0)
No comments yet. Be the first to share a thought on this story.
Request a correctionContact the newsroom