Hi,
has there been any thoughts about using tensorflow with Microsoft's DirectML (tf-dml currently in beta) to enable decent hardware acceleration with Intel and AMD cards in future?
In contrast to rocm which is a hassle to use and limited to Linux this is supposed to work on multiple OS and hardware.
No.
We already support PlaidML for AMD cards. Intel cards don't have enough Vram to run FS at any reasonable degree. We just can't justify the efforts to port to another backend.
While I of course understand and accept the reason, the performance of plaidML is pretty much disappointing (3.4 EGs on SAE with batch size 8 on 5700XT, not much better with vega64), so I had secretly been hoping on an improvement when tf-directml was announced, especially as Intel surely didn't hire Raja Koduri just for entry level gfx cards with 3gb vram. They hinted at starting with mass consumer market before stepping up.
Thank you for your response anyway 
AMD is significantly slower than Nvidia right now mostly due to software optimizations by Nvidia. That's not really PlaidML's fault and DirectML will almost definitely share in the slower speeds. On Intel, IF and WHEN Intel becomes viable for ML training we'll re-evaluate our decisions of course, but right now it's simply impossible to get good results from Intel GPUs and we're not going to do a bunch of speculative work which may not pay out with results.
AMD and other OSes are a constant thorn in our side.
At this moment in time, DirectML only supports Windows (and WSL). There may be plans for Linux in future, but I'm pretty sure that there are no plans to support macOS (could be wrong though).
We need to balance code maintainability and installation, with accessibility and features. On the face of it, DirectML has the potential to make our code more maintainable, however it creates further barriers to entry.
With Keras dropping multi backend support, we are looking at alternatives, but there really isn't a lot out there for AMD/macOS at the moment.
My word is final
