llama.cpp b9556 adds AMD RDNA3.5 GPU support for local LLMs
Enables running AI models on gfx1152/gfx1153 AMD GPUs with HIP.
The popular open-source LLM inference library llama.cpp has pushed release b9556, a significant update that adds HIP (Heterogeneous Interface for Portability) support for AMD's latest RDNA3.5 graphics architecture, specifically targeting the gfx1152 and gfx1153 GPUs. This means users with these new AMD GPUs can now run large language models locally with optimized performance through ROCm, AMD's compute stack.
Alongside the AMD addition, b9556 continues to support a broad range of platforms: macOS (Apple Silicon and Intel), Linux (x64/arm64 with CPU, Vulkan, ROCm 7.2, OpenVINO, and SYCL), Windows (x64/arm64 with CPU, CUDA 12/13, Vulkan, and HIP), and Android arm64. Some advanced backends like macOS KleidiAI and Linux SYCL are disabled in this release. The update also includes UI assets and various build fixes, maintaining llama.cpp's position as the go-to tool for on-device AI inference across consumer hardware.
- Adds native HIP support for AMD RDNA3.5 GPUs (gfx1152, gfx1153)
- Continues support for Apple Silicon, Intel, Linux, Windows, and Android builds
- Disables macOS KleidiAI and Linux SYCL backends in this release
Why It Matters
Expands local LLM inference to latest AMD GPUs, reducing cloud dependency for AI enthusiasts.