DirectML
A machine learning acceleration framework developed by Microsoft, built on DirectX 12. It enables AI inference and training acceleration using GPUs on Windows devices, including non-NVIDIA hardware, with broad support for AMD, Intel, and Qualcomm Adreno GPUs.
Timeline (last 90 days)
YoVoice's inference uses Apple Silicon MPS or Windows DirectML/CUDA acceleration, has no per-character billing, and audio data is not uploaded to third-party endpoints.
Microsoft can use DirectML to achieve hardware acceleration on a variety of GPUs, but actual performance varies considerably across graphics cards.
Local models require the operating system, drivers, and runtimes (such as DirectML and ONNX Runtime) to jointly perform graph partitioning and operator mapping in order to find the most suitable execution unit.
Windows' hardware-related optimizations are usually delivered in layers through the WDDM driver model, the DirectML runtime, and power management frameworks (such as Modern Standby and Energy Saver), and whether older platforms can benefit depends on whether the chip supports the same instruction sets and driver interfaces.
The phenomenon of slow first generation followed by faster subsequent generations stems from the GPU computation graph compilation and caching mechanism: the first run requires compiling the computation graph into low-level instructions adapted to the hardware, and after the result is cached it can be reused in subsequent runs.
对于AMD GPU用户，DirectML（微软基于DirectX 12的机器学习硬件抽象层）提供跨GPU厂商替代方案，但训练支持完整度不如CUDA成熟
ONNX Runtime的C++ API支持多种执行提供者，包括CUDA、TensorRT、DirectML，在Windows上与DirectML结合可利用各品牌GPU计算能力
All Facts (8)
AMD显卡可通过ROCm或DirectML方案运行Stable Diffusion，但兼容性和性能不如NVIDIA
85%UnverifiedYoVoice's inference uses Apple Silicon MPS or Windows DirectML/CUDA acceleration, has no per-character billing, and audio data is not uploaded to third-party endpoints.
50%UnverifiedMicrosoft can use DirectML to achieve hardware acceleration on a variety of GPUs, but actual performance varies considerably across graphics cards.
50%UnverifiedLocal models require the operating system, drivers, and runtimes (such as DirectML and ONNX Runtime) to jointly perform graph partitioning and operator mapping in order to find the most suitable execution unit.
50%UnverifiedWindows' hardware-related optimizations are usually delivered in layers through the WDDM driver model, the DirectML runtime, and power management frameworks (such as Modern Standby and Energy Saver), and whether older platforms can benefit depends on whether the chip supports the same instruction sets and driver interfaces.
50%UnverifiedThe phenomenon of slow first generation followed by faster subsequent generations stems from the GPU computation graph compilation and caching mechanism: the first run requires compiling the computation graph into low-level instructions adapted to the hardware, and after the result is cached it can be reused in subsequent runs.
50%Unverified对于AMD GPU用户，DirectML（微软基于DirectX 12的机器学习硬件抽象层）提供跨GPU厂商替代方案，但训练支持完整度不如CUDA成熟
50%UnverifiedONNX Runtime的C++ API支持多种执行提供者，包括CUDA、TensorRT、DirectML，在Windows上与DirectML结合可利用各品牌GPU计算能力
50%