I do not understand this take :) The vast majority of our code is vector length agnostic, as is required for SVE and RVV. GPU perf/TCO is not always better tha

I do not understand this take :) The vast majority of our code is vector length agnostic, as is required for SVE and RVV. GPU perf/TCO is not always better than CPU, if you can even get enough GPUs, and the latency is also a concern and inducement to do everything on the GPU, which can be limiting.