Nvidia DSX MaxLPS Boosts GPU Capacity by 40 Percent
Nvidia and Nscale demonstrated that DSX MaxLPS software can run 37% more GPUs under the same power budget, boosting aggregate throughput by nearly 50% for modern AI workloads.
Nvidia, in partnership with cloud provider Nscale, has evaluated its new DSX MaxLPS power-management software to address the issue of underused electrical capacity in AI data centers. Testing took place at Nscale's Verne campus in Keflavík, Iceland, running Kimi K2.5 workloads in FP4 on Nvidia GB300 NVL72 systems equipped with Blackwell Ultra GPUs. The evaluation compared a static baseline of 140 GPUs against a DSX MaxLPS-managed cluster of 192 GPUs, both operating under a strict 264.4 kW provisioned power budget. The software uses dynamic, policy-driven power allocation to distribute energy to active nodes rather than reserving peak power for idle ones.
The results showed a significant performance leap without exceeding the power limit. The DSX MaxLPS configuration increased normalized aggregate throughput by 49.2 percent, jumping from 1,084,503 tokens per second to 1,618,443 tokens per second. This improved efficiency from 4.10 to 6.12 tokens per second per watt. Meanwhile, individual instance performance remained steady: high-throughput instances went from 59,153 to 59,220 tokens per second, while low-latency instances held flat at 2,265 tokens per second. Total measured power rose from 166.2 kW to 198.9 kW, raising power-budget utilization from 62.9 percent to 75.2 percent, while mean GPU power grew from 97.0 kW to 131.8 kW.
For practitioners, this optimization introduces critical trade-offs. While median and P75 latencies remained within 5 percent of the baseline, the P99 time to first token increased by 17 percent from its 15.7-second baseline. This tail-latency spike highlights why operators must carefully evaluate performance metrics alongside raw capacity gains. To safely implement DSX MaxLPS, Nvidia outlines a five-step validation method: mapping the infrastructure boundaries, setting a clear baseline, rolling out policies cautiously, scaling up capacity with step-by-step testing, and establishing final production limits. Looking ahead, Nvidia plans to integrate these dynamic power-sharing capabilities into its future Vera Rubin NVL72 AI factories.
This is our own summary of reporting by NVIDIA Developer Blog
