Realtime AI News
Nvidia's Vera Rubin debuts in MLPerf with a training sweep and 10x agentic efficiency gain
Nvidia published Vera Rubin's first MLPerf results, claiming leadership across training, post-training and agentic inference, and a sweep of every MLPerf Training v6.1 benchmark. In its first MLPerf Endpoints submission, the platform reports more than 10x generational gains in throughput per watt for agentic AI.

Nvidia used its MLPerf debut for the Vera Rubin platform to make a full-stack claim: results published on November 16 show the new system leading across training, post-training and agentic inference.
The headline number is a sweep of MLPerf Training v6.1, where Nvidia says Vera Rubin took first place in every benchmark in the round, covering the full path from pre-training through post-training fine-tuning. For a company that treats benchmarks as product literature, the framing is part of the launch cadence rather than an afterthought.
The more consequential result sits in MLPerf Endpoints, a suite Vera Rubin entered for the first time. Nvidia reports more than 10x generational gains in throughput per watt for agentic AI, the metric that measures how much useful agentic work a system delivers for each unit of power.
That framing matters because agentic workloads behave nothing like single-shot inference. An agent may plan, call tools, read results and loop several times before answering, so the useful question becomes how many completed tasks a rack can sustain rather than how fast one token arrives.
By putting training, post-training and agentic inference on a single scorecard, Nvidia is arguing for platform-level capability rather than a peak number for one chip. That narrative carries weight with buyers whose cost structure is shifting from training runs toward always-on serving.
The caveat is the one that applies to every vendor submission. MLPerf permits engineering tuned specifically for the benchmarks, and the numbers describe idealised runs rather than a customer's production cluster; independent replication and real deployment data take longer to appear.
What to watch next is deployment velocity — when Vera Rubin capacity reaches cloud providers and large AI labs — and how the next MLPerf round looks once competing submissions land. Leaderboards describe a ceiling; purchasing decisions still turn on delivery dates, power budgets and software stack maturity.
Why it matters
The result shifts competition from per-chip peak performance toward agentic throughput per watt, which directly shapes the economics of large inference clusters.
Nearby Updates
All11/17, 00:00
Nvidia expands CUDA-X with cuPhoton, cuPotential and cuSpa for AI research agents
Nvidia expanded its CUDA-X libraries with cuPhoton, cuPotential and cuSpa, aimed at research in astrophysics, materials science and single-cell biology. The company says the speedups let scientists analyse larger datasets and model more complex systems, while giving AI agents callable tools for discovery.
11/17, 00:00
NVIDIA Extends Slurm to Put Quantum Processors Inside Supercomputing Workflows
NVIDIA is expanding Slurm's support for quantum processors so supercomputing centres can schedule quantum hardware alongside their existing scientific workloads. The move lands as nearly 30 centres prepare to deploy more than 50 quantum processors, a roughly sevenfold rise in deployments over five years.
11/17, 00:00
Nvidia unveils Rubin HPC, a 50,000-core-plus GPU built for science agents
At SC26, Nvidia announced Rubin HPC, a new GPU in its Vera Rubin platform aimed at simulation and data processing for science, engineering and analytics, with more than 50,000 CUDA cores on a single chip. Nvidia says the card scales data movement alongside compute, paired with sparse-data acceleration and seventh-generation NVLink, to shorten the loop from an agent's hypothesis to a simulated result.
11/17, 00:00
Lambda Adopts NVIDIA Photonics CPO Switches to Harden 10,000-GPU AI Factories
Lambda is deploying NVIDIA Photonics co-packaged optics (CPO) switches to connect NVIDIA GB300 NVL72 racks carrying more than 10,000 GPUs. The cloud provider is betting that a faster, more efficient and more resilient network is what keeps very large training clusters productive.