DSX MaxLPS Energy Efficiency: NVIDIA Reports Results

DSX MaxLPS energy efficiency was one of the topics NVIDIA presented on September 15 in a summary of Ian Buck’s keynote at the AI Infra Summit in Santa Clara. The company published Lambda’s result on Blackwell servers and also cited maximum values for suitable deployments of the Vera Rubin NVL72 platform. For AI data centers, these figures are relevant mainly because they relate to available power capacity limits.

The conference organizer confirmed that Ian Buck delivered a keynote on September 15 titled “Advancing Infrastructure for the Era of Agentic AI.” In its own material, NVIDIA describes the DSX MaxLPS system as an approach to power management during AI infrastructure operations.

DSX MaxLPS energy efficiency in the Lambda test

According to NVIDIA, Lambda operated 19 nodes on Blackwell servers within a power budget typically intended for 16 nodes. NVIDIA attributes to this result 24% higher token throughput and 23% better performance per watt.

This comparison concerns the capacity that can fit within a fixed power limit. In language-model inference, token throughput is one commonly used performance indicator, while performance per watt expresses the relationship between achieved throughput and energy consumption.

However, the published figures should be read as claims by NVIDIA and Lambda. No independent reproduction or complete methodology for the Lambda test was provided. The percentages therefore do not indicate a universal gain for every cluster, model, or workload.

Vera Rubin NVL72 and more GPUs within the same limit

NVIDIA also stated that DSX MaxLPS could enable up to 40% more GPUs without increasing the power limit in suitable deployments of the Vera Rubin NVL72 system. For the same framework, the company claims up to 35% higher token throughput.

NVIDIA introduced Vera Rubin DSX earlier; the current announcement focuses on newly presented energy-efficiency results, not the platform’s initial introduction. The stated maximums depend on configuration, workload type, and the conditions of a specific deployment. They therefore do not automatically mean the same improvement for all AI data centers.

For operators with a limited electrical connection or a fixed facility power limit, this aspect is particularly important. If the results were confirmed in production, a larger amount of computing capacity could fit within the existing power budget without immediately expanding the electrical supply.

Separate benchmark with Groq 3 LPX

NVIDIA’s technical blog also describes a combined Vera Rubin NVL72 and Groq 3 LPX result. For a specific scenario involving models with more than 2 trillion parameters, long context, and high interactivity, the company cites up to 35 times higher throughput per megawatt compared with GB200 NVL72.

This figure cannot be applied to all inference tasks. It is tied to the stated benchmark scenario and its model and operating conditions.

What will be important to monitor

To assess the presented figures, it will be important to see whether Lambda or NVIDIA publish a more complete methodology, cluster configuration, and data suitable for independent verification. Other relevant information will include real-world production deployments of DSX MaxLPS on Vera Rubin NVL72 systems and Vera Rubin benchmarks with Groq 3 LPX outside NVIDIA materials.

Power limits are among the practical constraints on expanding AI infrastructure. NVIDIA therefore presents performance per watt and throughput within a fixed power budget as important metrics, but the published maximums remain company results tied to specific conditions.

Verified and updated: 09/16/2026 06:27