A new fork of the open-jet megakernel, dubbed Geezuz, has been developed to significantly boost the performance of the Qwen3.8-27B model on dual RTX 3090 GPUs. This kernel architecture allows the entire forward pass to run as a single persistent CUDA kernel, with the fork specifically optimizing for two cards that sync and exchange data over PCIe. Benchmarks indicate that Geezuz achieves up to 190 tokens/sec for a single request and 622 tokens/sec for 8 concurrent requests, outperforming llama.cpp and approaching vLLM's performance. AI
IMPACT This kernel optimization could lead to faster local inference for large language models, potentially improving user experience and enabling new applications on consumer hardware.
RANK_REASON New kernel optimization for existing model and hardware.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →