NVIDIA DSX MaxLPS introduces policy-governed power sharing to optimize GPU use within a fixed power budget. It allows up to 40% more GPUs to operate at the same time. This approach monitors actual power consumption and reallocates power dynamically across resources.
A joint evaluation with Nscale tested Kimi K2.5 workloads on NVIDIA systems at Nscale’s Iceland data center. The test compared a static setup of 140 GPUs with a MaxLPS setup of 192 GPUs, both within the same power limit.
The results showed a 49.2% increase in overall throughput. The throughput per watt also improved significantly. Latency stayed within 5% of the baseline, but the time to first token at P99 increased by 17%, showing some tail latency impact.
Operators can validate MaxLPS using a five-step process. They define boundaries, set baselines, introduce policies carefully, add capacity gradually, and only set production limits after testing. This method helps ensure performance and safety.
MaxLPS combines hardware, thermal, and software controls. The control layer adjusts power limits based on real-time telemetry, maintaining electrical limits without increasing total power supply. This makes AI factories more efficient and flexible.
