How NVIDIA DSX MaxLPS Maximizes AI Factory Throughput and Efficiency

NVIDIA published a technical walkthrough of a joint evaluation with Nscale of NVIDIA DSX MaxLPS, a policy-governed power-sharing capability that dynamically allocates power across participating resources. According to the report, the approach enables customers to deploy up to 40% more GPUs within the same approved power budget.

The evaluation ran Kimi K2.5 workloads on NVIDIA GB300 NVL72 systems at Nscale's data center at the Verne campus in Keflavík, Iceland, powered entirely by renewable energy. It compared a static baseline of 140 GPUs, using 35 four-GPU nodes, against a DSX MaxLPS configuration of 192 GPUs, using 48 four-GPU nodes, under the same 264.4 kW provisioned power budget. The software stack included NVIDIA Dynamo and TensorRT LLM, with an 8K input sequence length and a 1K output sequence length.

DSX MaxLPS increased normalized aggregate throughput by 49.2%, and throughput per provisioned watt rose from 4.10 to 6.12 tokens/s/W. Per-instance throughput for high-throughput and low-latency workloads remained effectively unchanged at the displayed precision. Median and P75 latency stayed within 5% of baseline, while P99 time to first token increased 17% from a 15.7-second baseline, which the report says highlights the need to evaluate tail latency alongside capacity gains.

NVIDIA describes a five-stage validation process for operators: define the managed boundary, establish a representative baseline, introduce policies conservatively, add capacity incrementally with testing at each stage, and set production operating limits only once all objectives are met. The report states DSX MaxLPS combines dynamic power management with performance-per-watt techniques for future NVIDIA Vera Rubin NVL72 AI factories and infrastructure designed for 45°C liquid-cooling inlet operation, adding that Vera Rubin capacity projections should remain separate from the measured GB300 NVL72 evaluation.