# llama250.cpp Brings Vulkan and UMA Optimizations to AMD BC-250 Boards

> A specialized llama.cpp fork optimized for the AMD BC-250 mining board adds Vulkan scalar tuning, UMA mmap support, and cluster wire compression.

- Published: 2026-11-16T06:40:51Z
- Score: 7/10
- Tags: bc250, ai, pc
- Port / platforms: - -> Linux / Steam Deck, PC (Windows) (Tool)
- Web page: https://emu-radar.customfw.xyz/news/llama250-cpp-amd-bc250-tuning/
- Sources: https://github.com/Naster17/llama250.cpp/commit/f9f74b50a703d0983743dfd5ed0d86df3fbf74a4

Developer Naster17 released `llama250.cpp`, a fork of [llama.cpp](https://github.com/ggml-org/llama.cpp) tuned for the AMD BC-250 mining board powered by the Cyan Skillfish APU (`gfx1013`).

The `gfx1013` chip lacks cooperative matrix instructions and accelerated integer dot products, causing upstream [llama.cpp](https://github.com/ggml-org/llama.cpp) to use conservative Vulkan paths. This fork adds scalar-path optimizations, occupancy-aware tile selectors, and automatic `mmap` enablement when ReBAR is active. These changes boost prefill performance by up to 12% and decode speed by 2%. It also adds an optional BF16 wire compression mode for multi-board RPC clusters over 1 GbE.

To build and run on Alpine Linux with the Vulkan SDK:

1. Configure and build the project using CMake:
   `cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DGGML_VULKAN=ON -DGGML_NATIVE=ON -DGGML_RPC=ON` followed by `ninja -C build`.
2. Run inference with flash attention enabled and micro-batch size set to 512:
   `build/bin/llama-cli -m model.gguf -ngl 99 -fa 1 -ub 512 -t 6`.

Keep your KV cache in FP16, as `q8_0` measures slower on this hardware. Pair the code with the bc250rocm r3 kernel and Mesa 26.1 for expected performance, as older kernel versions remain flat or slower.
