Microsoft tests microfluidic cooling for next-gen AI chips

Microsoft is working on a new way to cool AI chips, using microfluidics to bring liquid directly inside the silicon. Unlike cold plates, which sit on top of the chip and are blocked by packaging layers, this approach carves tiny channels into the silicon itself. Coolant flows through these grooves and pulls heat away at the source.

In Microsoft’s tests, microfluidics removed heat up to three times more effectively than cold plates. It also cut the maximum GPU temperature rise by 65%, depending on the workload and chip type . AI models were used to map hot spots on the chip and direct the coolant with more precision.

Source: Microsoft

The company says this could make servers more compact, lower datacenter power use, and extend the life of GPUs and accelerators. With AI chips getting hotter every generation, Microsoft warns that cold plates may not be enough in the next five years. Work is now underway with partners to refine the technology, from packaging and coolant chemistry to full server integration.

I’m not sure if this technology will ever make its way into consumer chips, but since it’s basically fine-tuned by AI algorithms, at some point it will make no sense not to use it. I’m just wondering what will happen if any of the micro tubes gets stuck, will half of the GPU stop working? And cleaning must be a nightmare.

Source: Microsoft