AI BriefingiGPU boosts token speed by 10%↑
Key point
Moving the monitor to the iGPU raised llama.cpp token speed from 50t/s to 55t/s.
Details
While running Qwen3.6 35B A3B UD-IQ3_XXS (Unsloth) with llama.cpp on an RTX 4070 Super 12GB, connecting the monitor to the AMD iGPU and switching the primary display raised the token generation speed from 50 t/s to 55 t/s, an improvement of about 10%.
The key settings are as follows.
- Using
llama-server --fit on,--fit-target 256--flash-attn on--no-mmap,--mlock--no-context-shift--fit-ctx 262144,--predict 32768--cache-type-k q4_0,--cache-type-v q4_0--threads 8,--threads-batch 8
This is a case showing that by moving the monitor output to the iGPU and letting the PCIe GPU use more of its VRAM for inference, generation speed can be improved in some environments.
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
Apple Releases LensVLM-9B for Processing Compressed Text Images
Qwen Releases 'RecreationWorld' for Training AI Agents to Rebuild AppsQwen Intelligence Launches Three Mobile AI Agents and Open Benchmark Suite