How to switch GPUs between integrated graphic chip and independent adapter card?

[daoxian]

Hi there. I have a Windows computer with two graphic cards: one is an integrated intel graphic and the other is an nVidia graphic card. I know the way to set the running graphic card by nVidia driver & display tool application. But in fact, I want to control DirectML to use the one of better performance by my program codes instead of setting it manually. Is it possible?

[adtsai]

Absolutely! If you're calling into the DirectML native API directly, you can control this by controlling the device you supply to DMLCreateDevice. Specifically, when creating your D3D12 device you can specify an adapter to use. If you have multiple GPUs, you can automatically select the high-power device using IDXGIFactory6::EnumAdapterByGpuPreference with DXGI_GPU_PREFERENCE_HIGH_PERFORMANCE.

If using a higher-level framework like WinML, OnnxRuntime, or tensorflow-directml, there are also ways to control this:

• In WinML, use the LearningModelDeviceKind.DirectXHighPerformance option when creating your session.
• In OnnxRuntime, you can supply a custom device ID when initializing the DirectML EP using OrtSessionOptionsAppendExecutionProvider_DML
• In tensorflow-directml, you can control the GPUs visible to TensorFlow using the DML_VISIBLE_DEVICES environment variable.