How NVIDIA DSX MaxLPS Maximizes AI Factory Throughput and Efficiency | NVIDIA Technical Blog

This post details a joint NVIDIA and Nscale evaluation of DSX MaxLPS on NVIDIA GB300 NVL72 systems with Kimi K2.5 workloads. It demonstrates that DSX MaxLPS, using policy-governed power sharing, increased normalized aggregate throughput by 49.2% and throughput per provisioned watt from 4.10 to 6.12 tokens/s/W, while enabling 192 GPUs within the same 264.4 kW provisioned power budget compared to a static baseline of 140 GPUs. Per-instance throughput remained stable, and median/P75 latency stayed within 5% of baseline, though P99 time to first token increased by 17%. The post also outlines a five-stage validation process for operators to implement DSX MaxLPS.