Cooling heavily affects how reliably a chip operates and how quickly in ages. When you have a larger than average chip (say, the size of a dinner plate), you get a pretty tricky engineering problem. This is how we solved it at @cerebras . Historically, AI data centers, especially ones retro-fitted from the crypto mining era used air cooling. These are the extremely loud, higher energy consumption data-centers that are slowly being phased out. Water cooling, is a newer innovation that started at Cerebras and has since scaled to be implemented in NVIDIA and AMD datacenters as well. But is still tricky for such a large wafer. There are three things that we did. 1. Instead of the water flowing parallel across the chip (the chip is too big), water flows straight towards and away from the wafer 2. Each channel runs down close to the wafer without touching it, turns, and comes back out. Because every path is short, the outlet temperature is still close to the inlet temperature. 3. A custom cold plate to reduce thermal resistance between the wafer and the water. A lot of these are innovations that took many years to invent, by some of Cerebras' founders like JP, Sean Lie, and continue to improve.