AI and date centre roundup: Facing the energy challenge

Estimating AI energy consumption

Key for determining data centre connection requirements and optimising the allocation of workloads is the energy consumption of the processors and chips that make up the IT infrastructure. 

Such estimates are complex – the power consumption of a particular graphical processing unit (GPU) varies based on its configuration and the workload it is handling – and time consuming, taking hours or even days to complete with typical AI data centres having thousands of these GPUs performing operations to train and deploy AI models. 

But now with a new prediction tool, ‘EnergAIzer’, developed by researchers from MIT and the MIT-IBM Watson AI Lab, operators can determine within seconds how much power will be consumed by running a particular AI workload on a certain processor or AI accelerator chip. 

With this data centre operators could effectively allocate limited resources across multiple AI models and processors, improving their energy efficiency.  

In addition, the tool could allow algorithm developers and model providers to assess potential energy consumption of a new model before they deploy it. 

Tests of EnergAIzer reported using real AI workload information from actual GPUs found that it could estimate the power consumption with an error of about 8%, which is comparable to the traditional methods. 

“Because our estimation method is fast, convenient, and provides direct feedback, we hope it makes algorithm developers and data centre operators more likely to think about reducing energy consumption,” says Kyungmi Lee, an MIT postdoc and lead author of the publication on the tool. 

And the key to speeding up the prediction process? Like the use of AI in looking for patterns in data, the researchers found that AI workloads often have many repeatable patterns and that these patterns could be used to generate the information needed for reliable but quick power estimation. 

Also of interest
Meta looks to novel technologies to power AI data centres

A solid state transformer for AI data centres 

Enphase Energy has announced the development of the IQ solid state transformer for AI data centres as the industry moves toward higher density DC power architectures.  

The IQ SST is designed to replace traditional centralised power conversion with a distributed ‘supercluster’ architecture, with each 1.25MW IQ SST rack combining 342 intelligent semiconductor and software-defined power modules operating in a coordinated series/parallel configuration.  

The system is expected to deliver 98.5% efficiency and 99.999% availability through built-in redundancy, enabling continued operation with only 90% of the power modules participating. 

The IQ SST platform is designed without internal batteries and is intended to reduce or eliminate the need for rack level battery sidecars and traditional UPS systems in supported data centre configurations.  

For data centre operators, that means less power infrastructure with less floor space consumed by electrical equipment, and thus more space – as much as 50% more – available for GPUs. 

“AI is changing how power must be delivered to compute infrastructure. As AI racks move toward 800VDC architectures and megawatt-scale densities, we believe that distributed architecture is well suited to this transition, and it is what we are building,” says Badri Kothandaraman, president and CEO of Enphase Energy. 

The increasing rack power densities and loads that can swing from idle to full power and back several times per second are beyond what legacy architectures were designed for. Enphase’s new platform is designed to support both 800VDC and ±400VDC rack configurations defined by emerging AI data centre standards, with sub-millisecond response to dynamic AI loads. 

Full system demonstrations are expected late this year, with customer pilots in 2027 and volume shipments expected in 2028. 

AI data centres as grid assets 

A partnership between Nvidia, Emerald AI and Invenergy has been formed to power and advance a new class of AI data centres – or what the partners term ‘AI factories’ – that should be able to connect to the grid faster and operate as flexible energy assets. 

Based on Nvidia’s newly launched Vera Rubin DSX AI factory reference design, the next generation AI data centres can use co-located energy generation and storage as bridge power, then later harness these resources to flexibly supply the grid to support the broader power system. 

The DSX reference architecture can also support flexible AI data centres without co-located energy resources to achieve larger and faster power grid connections. 

Emerald AI’s Conductor platform orchestrates the computational flexibility alongside onsite generation, batteries and other behind-the-meter resources with the aim to deliver precise, grid-responsive power flexibility while ensuring quality of service for AI compute tenants. 

“AI factories are the engines of the intelligence era, and like any great engine, every system must be designed together – energy, compute, networking and cooling as one architecture,” says Jensen Huang, founder and CEO of Nvidia. 

“Nvidia and Emerald AI are working together to enable a future for AI where performance, efficiency and grid responsiveness can be tapped into immediately.” 

Power flexible AI data centres are estimated to be able to help unlock up to 100GW of capacity across the US power system by combining optimised infrastructure design with efficient use of existing assets and, where needed, new-build generation, while flexing during limited periods of grid stress to reduce the need for broader grid expansion to support reliability. 

Invenergy, along with AES, Constellation, NextEra Energy, Nscale Energy & Power and Vistra have committed to collaborating to evaluate optimised generation applications designed to power these next gen AI data centres, including through hybrid projects that use co-located power. 

Emerald AI and Nvidia report having trialed AI power flexibility demonstrations at five commercial data centres around the world over the last year.  

DSX Flex is expected to be deployed at commercial scale later this year at Nvidia's AI factory research centre in Virginia, planned as one of the world’s first power-flexible AI data centres with the Vera Rubin infrastructure.