NVIDIA is reportedly planning to switch to an entirely new cooling solution for the next-gen Rubin Ultra AI lineup, as it tries to mitigate the growing power requirements and thermal concerns.
NVIDIA Plans to Switch To Microchannel Cover Plates With Rubin Ultra, Squeezing Out Optimal Performance/Watt Levels
For those unaware, having an adequate cooling solution for next-generation systems is one of the top concerns for firms like NVIDIA, as power draw levels rise tremendously with each iteration, necessitating sufficient cooling mechanisms onboard. According to @QQ_Timmy, it is believed that NVIDIA is contacting cooling solution partners to integrate ‘direct-to-chip’ cooling through microchannel cold plates for its Rubin Ultra AI GPUs, which could prove to be a significant shift from conventional liquid cooling solutions and enable NVIDIA to achieve top-tier performance.
Let’s discuss microchannel cover plates (MCCP) mentioned here. It’s somewhat similar to direct-die cooling, which overclocking enthusiasts often employ with modern CPUs, and it involves a copper cold plate with microchannels embedded inside that allows coolant to flow around. This allows the plate to create local convection, reducing thermal resistance from the die to the fluid. It is similar to NVIDIA’s current conventional cooling solutions, except that the changes are made directly to the liquid-cooled plates on the chip, enabling significantly more efficient thermal management.
I could delve into more specifics about how MCCP works, but it’s essential to understand why NVIDIA requires this technology in the first place. Team Green is in a really tight product cadence, and scaling up from Blackwell to Rubin brings in much higher power requirements, especially in complex rack-scale systems, which is why it is necessary for NVIDIA to ‘think out of the box’. Even if the company doesn’t desire it, architectural advancements force them to invest in advanced cooling solutions.
It is claimed that NVIDIA has contacted Asia Vital Components, a Taiwanese thermal solution provider, to design the MCCP for Rubin Ultra. However, it was initially planned for Rubin; the tight timeline has given suppliers insufficient time to switch to an advanced solution. Interestingly, Microsoft recently announced microfluidic cooling, an approach similar to MCCP, with the key difference being a focus on ‘in-chip cooling,’ where the coolant is located either inside or on the backside of the silicon itself. This demonstrates that the industry definitely needs new cooling options.