Nvidia touts AI data centre software to boost output

Nvidia has outlined early deployment results for its DSX software and infrastructure platform for AI data centres, including utility demand-response activity and higher output within fixed power limits.

The announcement focuses on two approaches to power management for so-called AI factories: software that shifts computing workloads when the electricity grid is under strain, and software that reallocates power across servers to raise output without increasing a site's total energy budget.

One example came from Santa Clara, where Silicon Valley Power sent a signal to an AI facility to cut electricity use during a period of high demand. Emerald AI's Conductor software, which Nvidia described as an early example of the flexibility model behind DSX Flex, reduced the site's power draw from four megawatts to three while keeping higher-priority inference jobs running.

Nvidia said Silicon Valley Power has sent more than 200 demand signals to that facility, with each one triggering a successful automated response. The site is taking part in the utility's Flexible Load Interconnect Program, which treats AI facilities as resources that can lower demand when required.

Varun Sivaram, Chief Executive Officer of Emerald AI, described it as the first large-scale test. "We were watching with bated breath," Sivaram said.

Mansi Shah, Head of Product at Emerald AI, said the moment carried weight for the team. "This feels kind of like a SpaceX rocket launch," Shah said.

Power limits

Nvidia argues that electricity, rather than chips alone, is becoming the main constraint on AI data centre expansion. It framed output per megawatt as a key measure for operators trying to increase the amount of computing work completed within existing utility connections and site designs.

Its DSX MaxLPS software is designed to address that constraint by monitoring GPU and rack-level power use and shifting available headroom across systems according to workload demands. The aim is to reduce "stranded" power capacity left unused by static provisioning.

Cloud provider Lambda supplied the first deployment validation figures for DSX MaxLPS on Nvidia HGX B200 GPU servers. In tests on a five-rack, 19-node cluster, Lambda ran 19 nodes within the same facility power budget as 16 nodes at full power and recorded a 24% increase in cluster-wide token throughput, from about 4 million tokens a second to 5 million.

Lambda also reported a 23% improvement in performance per watt under that setup. According to Nvidia, the company serves more than 10,000 customers, ranging from AI-focused start-ups to larger cloud and computing groups.

Dave Ward, President of Cloud Services at Lambda, said the results suggested a way around fixed power ceilings. "With our proof of concept, we believe we've moved beyond the limitation of fixed power budgets," Ward said. "NVIDIA DSX MaxLPS paves the way to reclaiming stranded capacity and converting it into real-world usage, with significantly more compute density in the same footprint."

Broader design

DSX is not a single product but a broader set of software, simulation tools and reference designs covering operations, facility planning, resilience and power delivery. It also includes DSX OS for lifecycle management, DSX Sim for modelling facilities before deployment, and reference architectures spanning compute, networking, storage and building systems.

Nvidia also pointed to a longer-term shift in power architecture inside AI facilities. It is incorporating an 800-volt direct current design into DSX reference designs to reduce conversion complexity and support denser computing racks, with projected end-to-end efficiency gains of 3% to 5% compared with lower-voltage distribution approaches.

Nvidia argued that optimising isolated parts of a facility is no longer enough as AI systems scale. Faster chips can still be held back by networking limits, poor rack provisioning or cooling overhead, while liquid-cooled racks introduce further infrastructure demands that affect how much electricity reaches computing equipment.

For that reason, Nvidia is presenting DSX as a whole-facility approach rather than a chip-level efficiency tool. It argues that simulation before construction, software once systems are live and validated facility designs are all needed if operators want to increase useful work from each megawatt consumed.

The Santa Clara demand-response case is also intended to show that grid participation can move beyond theory. Nvidia said Emerald AI's Conductor responded to utility signals in under a minute, offering a model that DSX Flex is intended to generalise as the software develops.

Nvidia also said its projections show DSX MaxLPS could enable up to 40% more GPU capacity for Vera Rubin NVL72 AI facilities within the same megawatt budget in suitable deployment settings. "A one-gigawatt factory will never become a two-gigawatt factory," said Jensen Huang, Founder and Chief Executive Officer of Nvidia.