llama.cpp b9500 cuts Metal reset heartbeat from 500ms to 5ms
100x faster reset on Apple Silicon GPUs — tiny change, big impact.
Deep Dive
The latest llama.cpp release (b9500) reduces the Metal reset heartbeat from 500ms to 5ms. It also updates build targets for macOS, iOS, Linux, Windows, Android, and more.
Key Points
- Reduced Metal reset heartbeat from 500ms to 5ms, yielding 100x faster GPU resets on Apple Silicon.
- Builds now available for macOS, iOS, Linux, Windows, and Android with multiple backends (CPU, CUDA, Vulkan, Metal, ROCm).
- Open-source contribution from github-actions, signed with GPG key — standard for the llama.cpp project.
Why It Matters
Faster GPU resets mean smoother, lower-latency local LLM inference on Apple hardware, enabling more responsive AI apps.