AMD and Stability AI release Stable Diffusion for AMD NPUs
SDXL Base and Turbo models now run locally on AMD's AI Engine NPUs, enabling 1024×1024 image generation.
AMD and Stability AI have jointly launched Ryzen-AI SD Models, a set of Stable Diffusion models specifically converted to run on AMD's AI Engine Neural Processing Units (NPUs). This collaboration brings three key models to AMD's hardware: SD Turbo (based on Stable Diffusion 2.1 Distilled), SDXL Base, and SDXL Turbo, with Stability AI mirroring the releases. The initiative represents a strategic move to leverage NPUs—specialized chips designed for efficient AI inference—for local generative AI tasks, particularly image generation, without relying solely on powerful discrete GPUs.
Technically, the SDXL Base model is the most capable of the initial release, supporting 1024×1024 pixel image generation with its 2.6 billion parameters, closely matching the current ~3 billion parameter limit of consumer NPUs. While NPUs are less powerful than GPUs, they offer superior power efficiency for specific AI workloads, allowing them to handle tasks like real-time language translation or image generation while a GPU manages demanding applications like gaming. This development signals a future where local diffusion models could become ubiquitous on smartphones and laptops as NPU performance improves, potentially transforming applications like ComfyUI and making on-device AI generation a mainstream interaction.
- SDXL Base model enables 1024×1024 image generation with 2.6B parameters on NPUs
- NPUs offer power-efficient AI inference, complementing GPUs for multitasking (e.g., gaming + AI)
- Consumer NPUs currently support ~3B parameters, aligning with optimized model sizes
Why It Matters
Paves the way for efficient, local AI image generation on everyday devices, reducing cloud dependency.