Reduce GPU power limit: Reddit user finds performance sweet spot for local AI models
Adjusting MSI Afterburner settings on a gaming GPU boosts token rates without extra power draw.
A Reddit user, NotArticuno, conducted informal tests to see how reducing GPU power limits affected token processing and generation for an AI model. Using MSI Afterburner (a popular overclocking tool) on a gaming GPU, they ran Qwen 3.5:9b (likely a 9B-parameter variant of Alibaba's Qwen 3.5 model). The goal was to find if power limit reductions harm speed or can be compensated by memory clock boosts.
The results showed that lowering the power limit had minimal negative impact on token generation speeds. However, increasing the memory clock by 700–1000 MHz provided a moderate improvement in token generation across the board, even when the power limit was reduced. Core clock adjustments had very little effect. The user noted they did not test the memory clock increase at stock power limit, so the interaction remains partially unexplored. They also suggested that logging actual system power draw would be the next useful step to understand if core clock adjustments can simultaneously lower power consumption and improve performance.
- Reducing GPU power limit via MSI Afterburner did not significantly harm Qwen 3.5:9b token speeds on a gaming GPU.
- Memory clock increases of 700–1000 MHz moderately improved token generation despite lower power limits.
- Core clock adjustments had negligible impact, suggesting memory bandwidth is the bottleneck for local AI inference on consumer cards.
Why It Matters
Small, free tweaks to consumer GPUs can optimise local LLM performance, useful for hobbyists and professionals running AI on a budget.