NVidia's target market is GPU heavy workloads where the CPU is mostly for corrdinating the GPU-heavy work, moving data from IO to GPUs, etc. I've often seen in

NVidia's target market is GPU heavy workloads where the CPU is mostly for corrdinating the GPU-heavy work, moving data from IO to GPUs, etc. I've often seen in large-scale ML training that the CPU is mostly idling. Idle to the point where I've wondered if Presto/Spark workloads should be co-located on the idle cores for large $$$ savings.