At least in my deployment experience Arm is a core option now I think it competes very well with x86 and is often better in a lot of ways notably power use. IDK if you want to train on it but for ML inference and computational mathematics like fluid dynamics or areas like metal structure stress/strain dynamics it’s great and is a money saver whilst being effective still.

In cloud world Arm’s popularity is large. Google just announced alongside their TPU Ironwood release the release of N4A/C4A endpoints which are arm CPUs and Google are pushing both hard. AWS Graviton 3/4 instances have been great for low cost inference in my testing.

For local world I have actually been looking at making an Ampere Altra datacenter because it is close-ish to “local ARM xeon”. It would be similar to Gravitons in that it gives you Xeon-like performance with less power cost.

Since you are on Nvidia Grace you have an even stronger chip than Ampere Altra in terms of overall capabilities TBH, mostly because of the NVlinkC2C giving the fast connection between DRAM and VRAM. Its a better chip than any xeon or epyc because of that interconnection.