Repository navigation
MIGraphX EP #34187
|
Hello, I would like to ask regarding the MIGraphX EP on Windows. I am using Windows ML with DirectML as a vendor agnostic backend and it works. I wanna know regarding the AMD MIGraphX backend: Is it usable currently with Windows ML? Relevant issue microsoft/WindowsML#21. Can we confirm it is functional with an RDNA3+ GPU, so that I can get one and build my app's backend? Thanks! |
Replies: 2 comments 2 replies
|
Short answer: yes, it is listed as functional with RDNA3+ on Windows ML — but don't build a new backend on MIGraphX, it's deprecated. What the current Windows ML docs say (learn.microsoft.com, "Windows ML execution providers", last updated 2026-11-10):
So for your buy/build decision: an RDNA3+ card with a current driver meets the documented requirements, but for a new app backend use One note on why this is confusing: as the WindowsML#21 thread points out, the standalone ONNX Runtime MIGraphX EP documentation is Linux/ROCm-focused. The Windows story lives on the Windows ML execution providers page, not the ORT EP page. |
|
Dimitris (@kar-dim) Good question, and the analogy is not quite one-to-one. Per Microsoft's docs, For standard ONNX inference on RDNA 3+ it suffices, since this is AMD's supported path going forward (MIGraphX is deprecated in Windows ML 2.x and will be retired from the API in a new major release). If you depend on a specific MIGraphX graph pass, check AMD's docs for which backend gets selected for your workload. Source: https://learn.microsoft.com/en-us/windows/ai/new-windows-ml/supported-execution-providers (last updated 2026-11-10) |
Dimitris (@kar-dim) Good question, and the analogy is not quite one-to-one.
Per Microsoft's docs,
AMDGPUExecutionProvider"can select an AMD acceleration backend based on the available hardware and workload" — so it is a provider that picks a backend, not one fixed compute backend the way CUDA is. MIGraphX was the graph-compilation/optimization layer (the TensorRT analogue in your comparison), and AMD is folding that capability into the AMDGPU EP path rather than keeping MIGraphX as a standalone EP you select. It is not a ROCm backend — ROCm is AMD's Linux compute stack; on Windows the EP ships as an on-demand download via the ExecutionProviderCatalog APIs.For standard ONNX inference on …