Microsoft has presented its custom cooling solution for server and datacentre chips, featuring microfluid channels etched directly on the silicon to improve thermal transfer. The brand claims its design triples the cooling potential under specific conditions, opening the way for denser and higher-performing CPUs and GPUs to tackle the growing demand for AI.
Current cooling plates are currently reaching their heat removal limit, hindering potential chip designs, which need to consider aspects such as heat density and hotspots more than ever before. This is obviously a problem for all silicon-based processors, but particularly those used within the realms of AI.
To counter this, instead of going for a dual-phase solution like AMD, Microsoft has developed a cooling system that directly integrates microfluid channels into the chip’s silicon. These minuscule grooves – comparable to the diameter of a human hair – lie across heat-generating areas of the chip to remove heat more effectively. Since the coolant is in direct contact with the silicon, liquid temperatures don’t need to be as extreme to deliver similar cooling potential to traditional methods, saving energy and operational costs.
The team leveraged AI to help design these channels, customising them in a specific way to match the heat pattern of each chip. This results in a more efficient heat transfer, which in turn should allow for higher-density server arrangements as heat density becomes less of a concern. This should also open the way for more advanced architectures like 3D stacking, where heat removal is especially challenging.
Microsoft claims that its microfluidic design can dissipate more than 1kW/cm², which allows it to handle twice or triple the heat flux compared to standard cold plates. Additionally, the maximum temperature rise of the silicon is said to be reduced by 65%. While the focus of this innovation is on cooling the main chip, which can consume upwards of 900W in the case of the Nvidia GH200 Grace Hopper Superchip, HBM memory can also benefit. This should allow faster memory speeds for the most demanding tasks.
That said, these microfluid channels aren’t without their challenges. For instance, their extremely small size makes precision critical since each channel’s dimensions need to factor in the flow rate and mechanical strength. Not to mention that necessary changes in coolant chemistry to avoid gunk buildup or any issues that may block fluid circulation. With the absence of a large heat spreader to soak up the heat, any blockage could lead to catastrophic overheating.
Microsoft plans to explore integrating microfluidic cooling into its next-generation chips and is collaborating with its fabrication and silicon partners to scale this technology. The goal is to standardise it so the industry could bring its expertise and improve it.
