Method to Lock the NPU Frequency on Lunar Lake (LNL)

[Kepontry]

Hi, I'm benchmarking Qwen2-0.5B inference using GenAI on an LNL platform and observing up to 15% token generation latency variation between different runs, likely due to NPU frequency scaling (DVFS). For reproducible performance results, I need a way to lock the NPU to a fixed frequency.

Standard DVFS controls seem to act only as hints. I found a kernel patch for getting DPU frequency (link), but not for setting the NPU frequency.

Is there a recommended way to lock the frequency? As a potential workaround, would setting the pll_min_ratio and pll_max_ratio kernel parameters to the same be a viable approach?

Things I have done to minimize the variation:
sudo powerprofilesctl set performance
sudo sh -c "echo 1 > /sys/kernel/debug/accel/0000:00:0b.0/dvfs_mode"

[jwludzik]

Hi, @Kepontry

It is not possible to always use fixed NPU frequency. You can use pll_min_ratio and pll_max_ratio for testing purpose. Set it to max value, ex. pll_min_ratio=1000 pll_max_ratio=1000. Those values will be clamped to max allowed value for hardware by kernel driver. Still the frequency depends highly on DVFS algorithm that distributes power to CPU, GPU, NPU and other components. Setting pll_*_ratio provides higher chance for fixed NPU frequency

If you want to check how much power is provided to NPU, you can use "Intel® SoC Watch" tool for it

Setting CPU governor to performance is ok, it should bring some performance gain. You can try to use /dev/cpu_dma_latency to keep higher CPU states to get stable results