.. and then Nvidia brought their GPUs to the clusters and ate both Intel’s and AMDs lunch with a better version of VLIW. The CUDA programming model and hardware

.. and then Nvidia brought their GPUs to the clusters and ate both Intel’s and AMDs lunch with a better version of VLIW. The CUDA programming model and hardware IMO is quite successful at abstracting vectors the right way[TM], separating the problem of the grid setup from that of the kernel code (which you can still largely program in a scalar way if you’re not after the last ~30% of performance). IMO a shame that OpenCL never got good.