AI CPU resource use?

When using a CPU with Topaz products, do the various programs use single core or multi core? Or is threads? I ask because currently the only way I can use GigapixelAI compressed model without getting horrible artifacting is by using the CPU instead of the GPU. So are more cores/threads better or is single core performance more important?

I think the CPU is just able to calculate more precise. The GPU may lack of floating point precision.

The CPU is far less powerfull for this kind of calculations, so a different model is used.

The GPUs are capable of FP16, FP32, FP64.

Fp32 was the fastest amongst consumer cards until a few years ago
FP16 is picking up, most recent cards now support this at high speeds, higher than FP32
Fast FP64 is only available on enterprise cards or older Teslas and some older Kepler consumer.

Newer AMD Cards are very strong in FP16. Nvidia Cards with tensor cores also do FP16 fast and some nvidia consumer cards without tensor cores have dedicated FP16 cores (1660 for example).

Intel iGPUs are good at FP16, too…

So the transition will be towards FP16 - the precision is good enough for this kind of application.

1 Like