News
Local AI agents tuning the hardware they run on
Lower CPU boost clocks with no loss in inference speed, a GPU power limit the agent first called unfixable (+30–35% after the fix), auto-optimization on an RTX 5090 laptop and an RTX 3060 unlock. What this means for the Box.
A local AI system does not have to be tuned by hand. In our experiments, the agent running on local hardware analyses that same hardware, runs its own measurements and keeps only the settings that actually help. Below are four results from the last week, from newest to oldest.
Limiting CPU boost clocks: same speed, less heat
One underrated trick for running local AI is to limit CPU and/or GPU boost clocks. By voice, we asked the local AI agent to optimize the laptop that runs our main AI brain. It reduced the CPU frequency from 5.5 GHz to 4.5 GHz.
- LLM inference speed on Strata stayed exactly the same, with no loss.
- CPU temperature dropped by about 10%.
- CPU power consumption fell significantly.
- GPU temperatures dropped too, without limiting GPU clocks.
The CPU and GPU share a thermal budget, so reducing unnecessary CPU boost also helps the GPU. Nobody spent hours tuning: the agent was simply told to find the optimal limits, and it ran the experiments itself while work continued on other things.
Self-diagnostics: trust AI, but verify
Setup: Strata + Qwen3.8-Next-IQ3_S. The local AI declared a GPU power limit unfixable. We questioned the finding, cross-checked it externally (even the free version of ChatGPT is enough for this), fixed the limit and gained 30–35% performance.
The rule we apply: use AI for self-diagnostics, but cross-check critical findings.
Auto-optimization on an RTX 5090 laptop
Strata on a gaming laptop with an RTX 5090 (24 GB VRAM), model Qwen3.8-Flash-Next IQ3_S:
- Decode: 83.6 tok/s
- Agent decode: 99.8 tok/s
- Prefill 4K: 1,443 tok/s
- Prefill 16K: 2,132 tok/s
- First token: 0.28 s
After that, the agent optimized the same machine on its own:
- Warm decode: 63–110 → 109–118 tok/s
- Peak agent decode: 119.5 → 125.5 tok/s (+5%)
- The optimized configuration was saved, unstable settings were rolled back, and the server was shut down safely.
RTX 3060 12 GB: the unlock and the first gains
On one of our experimental computers (a small ITX case, 48 GB of RAM, an RTX 3060 12 GB) we enabled the NVIDIA unlock. The measured gain: up to +78% in GPU compute and +11% in AI inference prefill.
For perspective, with the next Strata work by coldniko the RTX 3060 12 GB + 48 GB RAM went from 32 to 60–80 sustained tokens per second. Our Mac Studio M3 Ultra (60 GPU cores, 256 GB RAM, Strata v0.2.6) delivered 61.4 tok/s decode and 734 tok/s prefill on the same Qwen3.8-Flash-Next IQ3_S model. The full results for this card are in a separate article.
What this means for a Box customer
- Settings are chosen by measurement, not by guesswork: the agent compares before and after and rolls back what does not help.
- Lower temperatures and power draw at the same speed mean quieter operation and less wear.
- Critical conclusions of the agent are cross-checked before they are acted on.
- Affordable hardware keeps getting faster with software updates alone.
Local AI optimizing the hardware that runs local AI.