NVIDIA Vera Rubin NVL72 claims 30x throughput gain for agentic AI workloads
NVIDIA announced significant efficiency advancements for its Vera Rubin NVL72 platform at the AI Infra Summit, emphasizing a shift from peak performance metrics to validated tokens per megawatt. The company highlighted that the new system delivers up to 30x higher throughput per megawatt compared to the previous GB300 NVL72 when running agentic coding sessions on the DeepSeek V4 Pro model. This…
Key points
- Vera Rubin NVL72 achieves 30x higher throughput per megawatt than GB300 NVL72 on agentic workloads.
- Lambda improved performance per watt by 23% using NVIDIA DSX MaxLPS power optimization software.
- Emerald AI demonstrated grid-responsive power management, allowing AI factories to throttle load based on demand.
Key partners demonstrated tangible results using NVIDIA’s DSX MaxLPS software, which optimizes power allocation across mixed workloads. Lambda reported a 23% improvement in performance per watt, achieving 5 million tokens per second on 19 nodes within the power budget of 16. Additionally, Emerald AI and Silicon Valley Power showcased a flexible-load program that allows AI factories to dynamically adjust energy consumption based on real-time grid signals, protecting priority workloads while supporting grid stability.
The event also featured benchmarks for the NVIDIA Vera CPU, with startups like Perplexity and ClickHouse reporting substantial latency and throughput gains for data-intensive and agentic tasks. NVIDIA positioned these updates as essential for scaling agentic AI, where long context windows and complex tool calls require infrastructure that maximizes output within strict power constraints.
The story so far
3 episodes →- NVIDIA Vera Rubin NVL72 claims 30x throughput gain for agentic AI workloadsthis story
AI Infra Summit: NVIDIA Vera Rubin and DSX Platform Advancements Showcase Energy Efficiencies of Optimizing Tokens Per Watt for AI Factories
NVIDIA Blog · 15 September 2026
Loading the full article…
This text was published by NVIDIA Blog and written by NVIDIA Writers. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
6sources- NVIDIA Vera Rubin NVL72 Posts First MLPerf Inference Preview ResultsPress · Unite.AI ·
- NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 DebutPrimary source · NVIDIA Blog ·
- CoreWeave Connects Hundreds of Rubin GPUs in One Multi-Rack ClusterPress · Unite.AI ·
- CoreWeave Brings Multi-Rack Nvidia Vera Rubin NVL72 Cluster Online — To Give Agentic AI Customers Greater Scale, Faster IterationPress · finance.yahoo.com ·
- Vera Rubin NVL72 Agentic Inference: 67x better Performance per DollarNewsletter · SemiAnalysis ·
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Hardware & Compute
All →- Google's Project Suncatcher satellite delivers first AI-generated haiku from orbit using Gemma 4 model on TPU · 5 src
- Powerlattice raises $86M to cut AI chip power use by over 50% · 1 src
- OpenAI CEO Sam Altman says home PC RAM is quite terrible for global allocation · 6 src
- Google’s cash burn drags down AI optics suppliers by 10-15% · 1 src
- Microsoft’s capex shortfall drags down CoreWeave, Nebius shares · 3 src
Comments
via GitHub Discussions