NVIDIA's DSX MaxLPS Boosts AI Factory Capacity by 40% Through Dynamic Power Sharing
Policy-governed power allocation increases throughput per watt by 49.2% in real-world tests
1Power sharing breakthrough
NVIDIA's DSX MaxLPS technology enables dynamic power allocation across GPU resources, allowing up to 40% more GPUs to operate within the same power budget. The system monitors actual consumption and reallocates unused capacity in real-time, addressing the common problem of infrastructure underutilization. In tests at Nscale's Iceland facility, this approach increased normalized aggregate throughput by 49.2% while maintaining the same 264.4 kW power ceiling.
2Real-world validation results
The joint evaluation used NVIDIA GB300 NVL72 systems running Kimi K2.5 workloads across 192 GPUs versus a 140-GPU baseline. Throughput per watt improved from 4.10 to 6.12 tokens/s/W, with median and P75 latency staying within 5% of baseline. However, P99 time to first token increased by 17%, demonstrating the need to balance capacity gains with tail latency considerations. The test environment used renewable energy at Nscale's Verne campus.
3Technical implementation
DSX MaxLPS combines chip, system, thermal, and software technologies to coordinate power allocation without increasing total site power. The system works within existing electrical limits from utility service down to individual GPUs. Dynamic Power Software provides the control layer for policy-governed allocation, enabling operators to safely reclaim stranded capacity that would otherwise remain unused due to conservative static power planning.
4Workload variability advantage
AI workloads naturally fluctuate between compute-intensive and idle states, creating power usage gaps that DSX MaxLPS exploits. Training cycles alternate between computation and synchronization, while inference workloads vary between prefill, decode, and network activity. Even identical models show power variance based on request shapes and concurrency levels. Traditional static allocation can't leverage these complementary power profiles.
5Validation methodology
Operators must follow a five-stage process to implement DSX MaxLPS: define managed boundaries, establish baselines, introduce policies conservatively, add capacity incrementally with testing, and set final operating limits. The evaluation confirmed these steps by measuring throughput, latency, interactivity, and power telemetry across four racks. Distributed workloads were confined to single racks to control for cross-rack variables.
6Performance trade-offs
While aggregate throughput increased significantly, the evaluation revealed important considerations. P99 latency increased by 17% despite stable median performance, highlighting the need to evaluate tail behavior. The optimal workload mix depends on complementary power profiles rather than simultaneous peaks. Operators must verify that added throughput doesn't compromise service quality or power compliance in production environments.
7Future infrastructure integration
For upcoming NVIDIA Vera Rubin NVL72 AI factories, DSX MaxLPS will combine with performance-per-watt techniques and 45°C liquid-cooling infrastructure. The technology provides a framework to turn natural workload variability into managed capacity, but physical infrastructure must still support the enabled capacity. Electrical distribution, cooling, and network fabric must be sized for lifecycle targets regardless of initial deployment scale.
8Operational transformation
DSX MaxLPS represents a shift from static to dynamic power management in AI factories. By making engineering trade-offs observable and controllable at fleet level, it allows operators to safely push infrastructure closer to theoretical limits. The technology is particularly valuable in power-constrained environments where physical expansion isn't possible, enabling more productive work within existing power budgets.