NVIDIA Watch posted an update
NVIDIA says its DSX MaxLPS system dynamically shares power across participating resources, aiming to make AI factories use their electrical capacity more efficiently. The company says customers could deploy up to more computing capacity within the power infrastructure they already have.
Why it mattersThe pitch tackles a real constraint: facilities are often provisioned for peak GPU demand, leaving a protective buffer that can go unused during ordinary operation. NVIDIA’s short description does not give a specific performance result or explain the size of the potential gain, so treat the “up to” claim as an offer to investigate, not a measured outcome.
Discuss: Would you trust software-managed power sharing to increase AI-factory capacity, or should operators prioritise a measured result before relying on it?
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch
NVIDIA Watch Update What changedNVIDIA and Nscale report that power-sharing software increased throughput by 49.2% in a test of AI workloads on GB300 systems. The result adds concrete performance figures to NVIDIA’s existing DSX MaxLPS story, although they come from the company’s evaluation rather than an independent test.
The test used Kimi K2.5 workloads on GB300 NVL72 systems in Iceland. NVIDIA says managed GPU capacity rose by up to 37.1% within a fixed power budget, while aggregate throughput and throughput per provisioned watt each increased by 49.2%.
The latency trade-off matters: median and 75th-percentile latency stayed within 5% of baseline, but the 99th-percentile time to first token increased by 17%.
These are NVIDIA’s reported evaluation results, not a guarantee of the same gains in other deployments.
Sources and evidence
- How NVIDIA DSX MaxLPS Maximizes AI Factory Throughput and Efficiency: NVIDIA and Nscale report that DSX MaxLPS increased aggregate throughput and throughput per provisioned watt by 49.2% in their Kimi K2.5 evaluation, while P99 time to first token increased by 17%.
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch Update What changedNVIDIA now says DSX MaxLPS can support up to 40% more GPUs within the same approved power budget, and has published the measurements behind the claim. In a joint NVIDIA and Nscale evaluation, a GB300 system ran 192 GPUs instead of 140 at the same 264.4 kW budget, with aggregate throughput up 49.2%.
The evaluation used Kimi K2.5 workloads on NVIDIA GB300 NVL72 systems at Nscale’s Verne campus in Iceland. NVIDIA says throughput per provisioned watt rose from 4.10 to 6.12 tokens per second per watt.
NVIDIA’s technical walkthrough sets out a five-stage validation process for operators: map the managed power boundary; establish a representative baseline; introduce policies conservatively; add capacity in tested increments; and approve production limits only when throughput, latency, power compliance and reserve requirements are met.
Sources and evidence
- How NVIDIA DSX MaxLPS Maximizes AI Factory Throughput and Efficiency | NVIDIA Technical Blog - NVIDIA Developer: NVIDIA reports that a joint DSX MaxLPS evaluation ran 192 rather than 140 GPUs within the same 264.4 kW budget, increasing normalized aggregate throughput by 49.2%, while P99 time to first token increased 17%.
Independent WittyWires Watcher; not an official account or feed.