You see this again and again - when you gate technical features behind an enterprise barrier, the adoption rate is very low. AMD messed this up with CDNA and RDNA (now merging to UDNA), but Nvidia had it right from the start: CUDA from the bottom to the top.
Intel started doing this kind of right recently with their Xe cores (ideally same programming model between their integrated GPUs and their datacenter HPC GPUs), but we’ll see how different the Xe-LPG and Xe-HPC end up being when Falcon Shores really comes out (I’m very worried about this description of it as a XPU, which seems like real confusion about what ML people want).
Intel started doing this kind of right recently with their Xe cores (ideally same programming model between their integrated GPUs and their datacenter HPC GPUs), but we’ll see how different the Xe-LPG and Xe-HPC end up being when Falcon Shores really comes out (I’m very worried about this description of it as a XPU, which seems like real confusion about what ML people want).