Post #1769688

@transcendentempress@eldritch.cafe Using the GPU tend to require a bunch of extra setup, as you are driving if from the CPU. And it's tricky to debug the code on it, so common practice is to write the code on the CPU first to make sure it works correctly. Meaning you are likely to write the code twice. GPUs tend to be easy to program when your computation doesn't need much synchronisation, but quickly starts to get nightmarish as more inter-thread dependency is added. Performance tend to not be portable between GPUs, especially of different vendors, because you need to tune the code to the architectural peculiarities and manage the buffers/bus differently. Long story short, it depends.