Why Are the Training Speeds Similar When Different Resource Flavors Are Used for Training on ModelArts Notebook Instances?

If your training job is single-process in code, the training speed is basically the same no matter when the notebook flavor of 8 vCPUs and 64 GB of memory or the flavor of 72 vCPUs and 512 GB of memory is used. For example, if your training job uses 2 vCPUs and 4 GB of memory, the training speed is similar no matter when you use the notebook flavor of 4 vCPUs and 8 GB of memory or the flavor of 8 vCPUs and 64 GB of memory.

If your training job is multi-process in code, the training speed backed by the notebook flavor of 72 vCPUs and 512 GB of memory is higher than that backed by the notebook flavor of 8 vCPUs and 64 GB of memory.