data center products are bottlenecked by chip on wafer on substrate packaging as they use chiplets.

Gaming GPUs do not- they have pricing power for consumer GPUs because they are a generation ahead of competitors.

model architecture changes not training objective (e.g self driving car).

Optical flow accelerator accelerates the computation of the direction of a moving image. It’s used as an input to ML, related in the same way a camera is to ML

because theyre measured with fp64 perf. not exactly a useful operation for deep learning.

80% market share says it absolutely is true. the pricing power is evident from all the things you listed. Radeon has not been able to keep pace with the advancements to computer graphics coming out of NVIDIA.