AMD 7000 Series GPU's

[MrHitman]

With the AMD 7000 series GPU's, I'm curious if Faceswap can be just as powerful on these AMD GPU's as NVIDIA's. I'm basically debating whether to go with the AMD 7800XT or the NVIDIA 4070. With my current card being limited at 8gb, I'm a bit worried of upgrading to only 12GB with the 4070.

Anyone able to help with these questions?

1. Can faceswap leverage AMD AI accelerators as it does with CUDA for NVIDIA? If so, is there a general rule of thumb on how much they compare in the 7000 series?
2. Can mixed precision be enabled for the AMD 7000 series?
3. Are there any other drawbacks besides the two above?

I've seen these drawbacks in the forum, but I haven't seen these revisited in a while so not sure how they apply to the 7000 series.

Thanks!

[torzdf]

Note, I have no experience with this, so these are just my opinion, they are not fact. Someone who has more experience may be able to help more.

1. Can faceswap leverage AMD AI accelerators as it does with CUDA for NVIDIA? If so, is there a general rule of thumb on how much they compare in the 7000 series?

Theoretically, yes. Currently Faceswap supports DirectML under Windows (probably won't be great performance wise) or ROCm on Linux. AMD recently released ROCm on Windows, so that may be possible to, but would currently require the end user to do some work, as we don't currently have an install path for ROCm on Windows.

Note: ROCm is a more involved install process. We can't pull in libraries for ROCm automatically, like we do for CUDA, so some manual work will be required on your part to get this running.

I would imagine (but cannot confirm) that ROCm will be slower than Cuda, but could be wrong.

I cannot confirm your other questions I'm afraid. I would have thought that AMD will have some kind of fp16 optimizations in place, but you would need to research this further.

[bryanlyon]

The AI accelerators in the 7000 series DO support FP16. However, we know almost nothing about their use. ROCm is supposed to support them, but we wont know how much it helps until we see them released. DirectML does not yet support them at all, and it's an open question if/when DirectML will add support for them.

That said, when comparing AMD's AI performance to Nvidia's you'll find that Nvidia's offerings are much more mature, stable, and faster. AMD claims that with the accelerators they can achieve 12 tflops, which is fraction of the 4070's claimed 29.77 TFLOPS but we'll have to see if the software can actually achieve that 12 tflops in actual use.

There is just no way that AMD can compete using 1st generation accelerators. It's a good step, but it's still just their first step. But if you only care about AI speed, Nvidia is undoubtedly the better option still.