Nvidia develops technology to help AI data centres ease pressure on power grids

Nvidia is developing technology that could allow large AI data centres to automatically reduce their electricity consumption when power grids come under pressure.

The chipmaker has been showcasing its DSX platform as the technology industry searches for ways to accommodate rapidly increasing demand for electricity from artificial intelligence infrastructure.

One component, DSX Flex, is designed to allow AI facilities to respond to signals from electricity networks and temporarily reduce power consumption without shutting down their most important computing tasks.

Lower-priority AI workloads could be slowed or paused when electricity demand is particularly high before automatically returning to normal when additional capacity becomes available.

The approach could potentially address one of the biggest constraints facing the continued expansion of artificial intelligence: access to sufficient electricity.

Building new power stations, transmission lines and other grid infrastructure can take years, while the rapid construction of increasingly large AI data centres is creating demand for substantial amounts of additional capacity.

Nvidia argues that making data centres more flexible could allow existing electricity infrastructure to be used more efficiently.

The technology has been developed alongside work involving energy software company Emerald AI and Silicon Valley Power, the municipally owned electricity provider serving Santa Clara in California.

A recent commercial-scale demonstration used Emerald AI’s Conductor software at an Nvidia AI facility.

When Silicon Valley Power requested a reduction in electricity consumption, the system automatically adjusted flexible computing workloads in less than a minute while protecting higher-priority AI work.

Rather than requiring every data centre to operate continuously at its maximum potential electricity demand, utilities could therefore treat some computing activity as flexible.

That could become important when deciding whether there is sufficient grid capacity to connect new AI infrastructure.

Electricity providers traditionally need to consider whether their networks can supply a large customer even during periods when overall demand is particularly high.

If a data centre can reliably demonstrate that it will reduce its consumption during those periods, utilities could potentially allow more computing capacity to connect to existing infrastructure.

Nvidia is seeking to build that capability into DSX Flex.

The system is designed to receive information including requests to reduce electricity consumption, demand-response events and changes in electricity prices.

It can then respond according to a predefined hierarchy of workloads, keeping critical computing jobs operating while temporarily reducing less urgent work.

Another possibility is shifting suitable AI workloads between data centres.

If electricity supplies are constrained in one region while another facility has spare capacity, some computing activity could potentially be moved rather than simply stopped.

The technology could have implications well beyond individual data centre operators.

Electricity consumption associated with AI is becoming an increasingly important issue for governments, utilities and communities as companies invest billions in new computing infrastructure.

The challenge is particularly acute because large AI facilities can require hundreds of megawatts of electricity and may be concentrated in areas where grid capacity is already constrained.

Nvidia and Emerald AI have also participated in testing in the UK.

During a demonstration involving National Grid and an AI facility near London, a cluster of Nvidia Blackwell Ultra GPUs was able to reduce electricity demand substantially in response to simulated grid events while maintaining priority workloads.

Nvidia’s longer-term aim is to make this flexibility a standard part of the infrastructure used to construct what it calls “AI factories”.

The first dedicated commercial deployment of DSX Flex is planned for Nvidia’s AI Factory Research Center in Manassas, Virginia.

The 96-megawatt facility is expected to test grid-responsive computing at a much larger scale.

There is also a clear commercial incentive for Nvidia.

Access to electricity is becoming a potential constraint on how quickly customers can deploy the huge numbers of GPUs required for advanced AI systems.

Finding ways for data centres to obtain grid connections more quickly could therefore remove an important obstacle to further expansion of AI infrastructure.

Whether flexible computing can deliver those benefits across electricity networks at much larger scale remains to be demonstrated.

But the technology represents an increasingly important shift in thinking about AI’s energy requirements: instead of treating data centres purely as enormous permanent electricity loads, their computing workloads could increasingly be managed according to the amount of power available on the grid.

Open article on Cheshire Today