Comment by jaxytee

Running large MOE models won't saturate the EPYC memory.

You'll never see that 200 Gb/s running inference.