We all like GPUs. Our author thinks they’re sufficient reason to stick to Linux. They’re not liable to go away, and continue to grow in power and importance. And a shiny new A100 can slurp a gigabyte of data a second, I think it was? You can buy eight of them in a single server.

But the rule of thumb they tell me is: If you try to use CPUs to move data to all these GPUs, your performance is going to tank. If at all possible, you want tech like RDMA, something that throws bits directly from your filesystems to your GPUs. This will only become more important as GPUs get more capable and data hungry.

Can Cthulix actually control enough I/O to build the future of distributed computing? The approach described wants to explicitly minimize filesystem use in favor of message passing. I’m not sure how that’s supposed to be implemented exactly? but there are only a few ways to get the I/O performance you want for the big heavy GPU workloads, and this doesn’t naively sound like it’s one of them, but maybe that’s because it’s a big question mark.