On-Prem

Systems

Infineon promises 12kW PSUs for next generation of power-hungry AI servers

How else are you going to feed all those 1,000-plus watt GPUs?


Not long ago an entire rack of systems might consume 10kW or less of power. Today, that's barely enough to run a single GPU server, and those systems are only expected to get hotter and greedier for power amid the AI boom.

To keep those GPUs adequately fed, Infineon is developing new, high-efficiency power supply units (PSUs) capable of delivering up to 12kW of energy to the system, up from around 3.3kW today.

To do this, Infineon says it's using a combination of silicon, silicon-carbide, and gallium-nitride technology. You may recognize gallium nitride, or GaN, as the material that's helped to shrink consumer chargers in recent years.

Silicon-carbide (SiC) on the other hand, has gained widespread attention amid the proliferation of electric vehicles. Generally speaking, SiC is more efficient at higher voltages, making it a prime choice for the traction and DC-to-DC inverters commonly found in EVs. However, as Infineon notes, that technology also makes sense in high-wattage PSUs needed to fuel hotter and hotter GPUs and AI accelerators.

The German giant says its next-gen PSUs will be capable of achieving 97.5 percent efficiency, when they start shipping early next year. Infineon contends these will help to drive down carbon emissions and power losses.

The first of these PSUs will be an 8kW unit that Infineon says will be able to support AI racks up to 300kW or more. Meanwhile, its 12kW power supplies are set to follow sometime later.

The power demands of modern AI systems have proven to be quite the headache for datacenter operators. For instance, as The Next Platform previously reported, Digital Realty has faced challenges with power as it worked to support larger deployments of Nvidia's DGX platform, which packs four 10.2kW servers into a single rack.

But while eight and even 12kW power supplies will no doubt simplify designs by reducing the number of PSUs required, power is likely to remain an ongoing challenge. As Schneider Electric wrote in a report last summer, PSUs aren't the only limiting factor, we're also going to need beefier power distribution units (PDU).

According to Schneider, at 120/208 volts it would require five 60 amp circuits to power an 80kW rack with single failure redundancy. This, they note, poses a variety of challenges and could impede airflow. Because of this, they recommended switching to 240/415 volt PDUs, and even at these voltages, Schneider warned this may not be enough.

But even if you can sort out rack power delivery, that doesn't do you much good if your local utility isn't up to the task. Amazon is reportedly rationing resources at its Dublin datacenters due to power constraints, a factor that previously led Microsoft to build a 170 megawatt gas plant to even out disruptions to the grid.

Meanwhile, in the UK, National Grid CEO John Pettigrew warned in March that datacenter power consumption could rise by as much as 500 percent over the next decade and would require major infrastructure improvements to support that growth. ®

Send us news
11 Comments

So much for green Google ... Emissions up 48% since 2019

AI datacenters blamed for the increase, even as Chocolate Factory bets on AI to fix it

AI to boost datacenter capex by 28.5% and become the top server workload

Cooling tech left to sweat the details

AMD buys developer Silo AI in bid to match Nvidia's product range

First it comes for market leader's GPUs ... now it's nibbling at software

OpenAI develops AI model to critique its AI models

When your chatbots outshine their human trainers, you could pay for expertise ... or just augment your crowdsourced workforce

64% of people not happy about idea of AI-generated customer service

Not unreasonably, nearly half worried it would give them the 'wrong answers'

Cloudflare debuts one-click nuke of web-scraping AI

Take that for ignoring robots.txt!

A friendly guide to local AI image gen with Stable Diffusion and Automatic1111

A picture is worth a 1,000 words... or was that a 1,000 TOPS

Palantir, Oracle cosy up to offer Karp firm's tech across Big Red's cloud

Foundry and AI plaform available in OCI

China’s homebrew openKylin OS creates a cut for AI PCs

Devs of OS named for a mythical beast join in the 'local models will will deliver legendary productivity' trope

Babel fish? We're getting there. Reg reviews the Timekettle X1 AI Interpreter Hub

A handy standalone translator, but you'll need deep pockets, both figuratively and literally, if you want one

Meta training AI models on citizen data gets a hard não from Brazil

Zuckerborg's justification isn't good enough, says watchdog

A friendly guide to containerization for AI work

Save the headaches, ship your dependencies