I kind of hope so. AMD and Intel hasn't really had a tight software department. For example, offloading some math to the Intel GPU has been in the OpenMP 4.5 spec since 2015 in a really easy way. It is supported by using Intel® oneAPI Base Toolkit AND the Intel® oneAPI HPC Toolkit. Which... no one uses.
That we still don't have any good GPGPU resources is just crazy to me.
Heterogenous computing with TPUs/GPUs/DSPs and other chips should be standard by now.
>Heterogenous computing with TPUs/GPUs/DSPs and other chips should be standard by now.
It sounds nice in theory but in practice is hard. Writing CUDA or OpenCL is not exactly pleasant or easy and compilers do a poor job at vectorizing code.
Se we use accelerators when it's an absolute must.
So we also need to be able to see the IR/ASM. It's not like the CPU only compilers are great at it either, quite a lot of handholding is needed there too, but one step at a time.
That we still don't have any good GPGPU resources is just crazy to me.
Heterogenous computing with TPUs/GPUs/DSPs and other chips should be standard by now.