NVIDIA power software lifts token throughput in first Lambda validation

NVIDIA has published the first deployment validation of DSX MaxLPS, the power management layer of its DSX platform for AI data centres. The test was run by GPU cloud provider Lambda and the numbers were released at the AI Infra Summit.

Lambda ran the software on a five-rack, 19-node cluster of NVIDIA HGX B200 servers. The 19 nodes were held inside the same power budget as 16 nodes running at full power. Cluster-wide token throughput rose 24%, from roughly 4 million to 5 million tokens per second, while performance per watt improved 23%.

DSX MaxLPS monitors GPU and rack-level consumption and reallocates headroom between nodes, recovering capacity that static provisioning leaves stranded. Training and inference draw power differently, so mixed workloads benefit most. NVIDIA projects that in suitable environments the approach can enable up to 40% more GPU capacity for next-generation Vera Rubin NVL72 factories within the same megawatt budget.

Data centres that flex with the grid

The second strand is grid participation. NVIDIA's Eos AI factory runs Emerald AI's Conductor platform as a participant in Silicon Valley Power's Flexible Load Interconnect Program. When the grid is constrained, low-priority jobs are slowed or rescheduled while priority inference keeps running.

During an August test the facility's draw fell from four megawatts to three with no operator involved. Emerald AI's Varun Sivaram said it was the first time the company had deployed across thousands of NVIDIA GPUs. Silicon Valley Power has since sent more than 200 demand signals to the site, with Conductor responding in under a minute each time.

The first dedicated commercial DSX Flex deployment will be a 96-megawatt Vera Rubin facility in Manassas, Virginia, at NVIDIA's AI Factory Research Center. DSX was introduced at GTC Taipei in May.

Full details are in the NVIDIA Blog post.