2 ms·
>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
by DeathArrow 7y ago
>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.
- danielscrubs 7y agoSo 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.