DeepFaceLab
DeepFaceLab copied to clipboard
iperov
Does an Intel arc a750 or 770 work on deepfacelab?
As the title states does this program work with an Intel arc a750 or 770. I was going to upgrade my computer so I can move over to saehd, and get much faster iterations. Whilst looking at graphics cards I noticed the A770 was cheaper and performed better than an RTX 3060. But I read a couple reviews saying that it doesn't work well with deep learning programs. So I am asking if it would work with deepfacelab because it does work on some.
Yes it does work with DirectML version of DFL. However the existing bundle of DirectML is using an older version of Tensorflow, which has a poorer performance.
I've filed a PR yesterday to migrate DFL to DirectML with Tensorflow 2. It provides a much better performance, which in my benchmarks on A750 is 45% ~ 133% faster, depending on the model.
Unfortunately it has been rejected since DFL is no longer maintained from what it seems, but I have provided instructions to try it out: https://github.com/iperov/DeepFaceLab/pull/5627
update: though I don't have a 3060 to test on, so can't tell you whether DirectML is faster than the CUDA version.
Yes it does work with DirectML version of DFL. However the existing bundle of DirectML is using an older version of Tensorflow, which has a poorer performance.
I've filed a PR yesterday to migrate DFL to DirectML with Tensorflow 2. It provides a much better performance, which in my benchmarks on A750 is 45% ~ 133% faster, depending on the model.
Unfortunately it has been rejected since DFL is no longer maintained from what it seems, but I have provided instructions to try it out: #5627
update: though I don't have a 3060 to test on, so can't tell you whether DirectML is faster than the CUDA version.
I'll try both on my 4090 and report back which version is quicker.
Your version - ~550ms +/- 10 - AMP model res 384 dims 128
CUDA version - ~604 + 100 - AMP model res 384 dims 128
Directx12 V - ~570 + 50 - AMP model res 384 dims 128
Your version - ~190 - AMP model res 320 dims 80
CUDA version - ~270 - AMP model res 320 dims 80
Directx12 V - ~200 - AMP model res 320 dims 80
Your version - ~450ms - saehed ud model res 384 dims 88
CUDA version - ~477 - saehed ud model res 384 dims 88
Directx12 V - ~495+ 20 - saehed ud model res 384 dims 88
Generally yours seems faster than even the cuda version.
Issue solved / already answered (or it seems like user error), please close it.
