NVIDIA Announces DSX MaxLPS to Boost AI Factory Efficiency

NVIDIA has announced the release of DSX MaxLPS, a technology designed to maximize AI factory throughput and efficiency by dynamically allocating power across participating resources. According to the announcement, DSX MaxLPS uses policy-governed power sharing to enable up to 40% more GPUs within the same approved power budget.

A joint evaluation between NVIDIA and Nscale demonstrated the effectiveness of DSX MaxLPS. The evaluation ran Kimi K2.5 workloads on NVIDIA GB300 NVL72 systems at Nscale's Verne campus in Keflavík, Iceland. The comparison was made between a static baseline of 140 GPUs and a DSX MaxLPS configuration of 192 GPUs, both under the same 264.4 kW provisioned power budget. The results showed a 49.2% increase in normalized aggregate throughput and a rise in throughput per provisioned watt from 4.10 to 6.12 tokens/s/W.

While per-instance throughput for high-throughput and low-latency workloads remained unchanged, there was a 17% increase in P99 time to first token, indicating the importance of evaluating tail latency alongside capacity gains. Median and P75 latency stayed within 5% of the baseline.

Operators are advised to follow a five-stage process to validate DSX MaxLPS: 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 when all objectives are met.

NVIDIA DSX MaxLPS is designed to reclaim stranded capacity in power-constrained AI factories and can be applied across diverse environments. The technology combines dynamic power management with performance-per-watt techniques and infrastructure designed for 45°C liquid-cooling inlet operation in future NVIDIA Vera Rubin NVL72 AI factories.

Source: nvidia-developer