GFX1151 Engine
An open-source inference engine developed by Bilibili content creator "防暴键盘," optimized specifically for the AMD AI Max+ 395 platform. It supports local deployment of large language models and provides targeted optimizations for Mixture-of-Experts (MoE) architectures as well as multiple weight quantization formats.
Timeline (last 90 days)
GFX1151 Engine is an inference engine written by the UP host themselves, supports Windows, and has been open-sourced on GitHub.
GFX1151 Engine's decoding (conversational generation) speed on the AMD 395 is 30 to 40 tokens per second, depending on the acceptance rate of speculative sampling.
The framework's prefill speed cap is approximately 900 to 1,000 tokens, and generally hovers around 900 on Windows after resource contention.
The closed-source Hologen Flash Server (Linux) can achieve a prefill speed of 1,200 to 1,400 tokens, outperforming GFX1151 Engine.
Because VRAM is fully occupied, using the machine simultaneously as a regular Windows workstation significantly degrades inference performance, making it unsuitable as a productivity tool.
All Facts (5)
GFX1151 Engine's decoding (conversational generation) speed on the AMD 395 is 30 to 40 tokens per second, depending on the acceptance rate of speculative sampling.
50%UnverifiedThe framework's prefill speed cap is approximately 900 to 1,000 tokens, and generally hovers around 900 on Windows after resource contention.
50%UnverifiedThe closed-source Hologen Flash Server (Linux) can achieve a prefill speed of 1,200 to 1,400 tokens, outperforming GFX1151 Engine.
50%UnverifiedBecause VRAM is fully occupied, using the machine simultaneously as a regular Windows workstation significantly degrades inference performance, making it unsuitable as a productivity tool.
50%UnverifiedGFX1151 Engine is an inference engine written by the UP host themselves, supports Windows, and has been open-sourced on GitHub.
50%