AI BriefingParallel Prefill with ANE+GPU Implemented on Apple Silicon
Key point
A new technique leveraging parallel ANE and GPU execution on Apple Silicon improves LLM Prefill speed by up to 50%.
Details
A technique that simultaneously utilizes the ANE (Neural Engine) and GPU in Apple Silicon environments to enhance LLM Prefill speed has been applied to the omlx project.
Previously, performance degradation occurred when using the ANE due to synchronization issues. However, by optimizing through a method that shards only parts of the MLP and GDN, a ~50% improvement in Prefill speed was achieved on the Qwen3.8 27B q4 model based on the M3 Ultra.
- In an M1 Pro (32GB) environment, a performance improvement of approximately 19% was confirmed for the 9B model, increasing from 280 -> 334 tokens/sec.
- This feature requires a custom kernel (
OMLX_WITH_CUSTOM_KERNEL=1), and note that peak memory usage increases by approximately 2x.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Underdog Releases 'Husky', a Model-Specific Inference Engine Up to 4.5x Faster than MLX
KV Cache Research: Limitations of LRU Alternatives and Failure in Capacity-bound Environments