I would like to know why they don't use GPUs for training but vCPUs.