Device support

[RoySegal]

Hello,

I would like to learn how to add a new proprietary AI processor capabilities under DirectML.

As far as I understand, the current support is for Direct3D, DirectX 12 devices, which means GPUs only.

DirectML official documentation is stating:

DirectML requires a DirectX 12 capable device. Almost all commercially-available graphics cards released in the last several years support DirectX 12.

I also found the following in the official documentation of WindowsML:

Windows ML (WinML) is a high-performance, reliable API for deploying hardware-accelerated ML inferences on Windows devices. DirectML provides the GPU backend for Windows ML.

However, my intention is to add a new AI processor device (and not a display device), which I want to support in order for others to be able to use WindowsML for NN deployment over my AI processor.

Does implementing DirectML API is the right way to go? If so, can you please direct me to the relevant APIs?

If not, what is the right layer I should interface to?

Many thanks, Roy

[jstoecker]

The interface layers are DirectML -> Direct3D 12 -> WDDM/MCDM -> Hardware Driver -> Hardware. Supporting a new DirectX-visible hardware device means providing drivers that implement either WDDM or MCDM:

• DirectX devices (or adapters) were, for the longest time, essentially synonymous with render devices or GPUs; this is why you see most of the docs referring to GPUs and graphics cards. The traditional approach is to implement the Windows Display Driver Model (WDDM), which (as the name suggests) involves handling a bunch of display-related functionality.
• There is also a newer Microsoft Compute Driver Model (MCDM), which enables compute-only hardware accelerators (like AI processors) to work with the DirectX stack.

If you are able to implement MCDM for your AI processor, then DirectML should simply work out of the box. Of course, custom silicon often means unique constraints, and I don't want to trivialize the effort required here. Implementing a compliant driver is not a small task, so I suggest e-mailing us askdirectml@microsoft.com if you want to discuss this in more depth!

[RoySegal]

Thank you for the quick and elaborate reply. I will definitely look into MCDM which is much more suitable for this usecase. To your knowledge - are there any non-GPU devices currently supported by DirectML? Alternatively - are these devices using the MCDM?

Again, your help is appreciated.

[jstoecker]

Yes, one example of a non-GPU device that works with DirectML is "Intel’s new neural VPU - a dedicated, power efficient hardware block focused on AI acceleration at low power which is integrated into the upcoming Meteor Lake platform." See https://community.intel.com/t5/Blogs/Tech-Innovation/Client/New-AI-Engine-Delivers-Power-Efficient-AI-with-DirectML/post/1489703.