Capping CPU Power Cut Laptop Temperatures from 96 to 68 °C Without Slowing Down an LLM

A local language model on a laptop can run fast, but the price of that speed sometimes looks odd: the processor gets scorching hot and the fans drown out your voice, even though the graphics card carries most of the computational load. The author of the experiment ran into exactly this on a Lenovo ThinkPad P16 Gen 2 with an RTX 2000 Ada and 8 GB of video memory.

He ran a 12-billion-parameter Gemma model in a 4-bit QAT variant in LM Studio. Together with an 8192-token context, it reportedly fit in VRAM and generated about 12 tokens per second. But the CPU heated up to 95–96 °C, and the cooling system switched to a very noisy mode.

Why Limiting the CPU Did Not Slow the Model Down

The logic of the experiment is simple. If the model weights and working context sit in video memory, the main math is done on the GPU. The processor keeps serving the application, data preparation and other system operations, but its maximum frequency may not affect token output speed.

To test this, the author created a separate AI profile in ThrottleStop. In it he disabled CPU turbo boost and changed the temperature threshold via PROCHOT Offset. By his estimate, processor performance dropped by roughly 40–50% relative to peak mode.

The result was telling: the RTX 2000 Ada kept producing the same 12 tokens per second. In other words, in this particular configuration the processor was not the bottleneck of generation. The freed-up thermal headroom let the overall cooling system run more calmly.

What Changed After Tuning the Cooling

Capping the frequency alone was not enough for the author. Using TPFanControl, he also set up earlier and smoother speed switching for the two fans. Instead of abrupt jumps in RPM, the cooling began reacting in advance to rising temperature.

According to the author's measurements, peak CPU temperature fell from 95–96 to 68 °C. The graphics card, meanwhile, heated up to no more than 59 °C, and fan noise turned from a noticeable drone into background rustling. Generation speed stayed at 12 tokens per second.

The practical meaning here goes beyond a single ThinkPad. Before buying a more powerful computer, it is worth identifying the real bottleneck: VRAM capacity, memory bandwidth, GPU, CPU, or offloading part of the model into system RAM. If the model runs entirely on the graphics card, extra processor power may mostly yield heat and noise.

Where the Experiment Should Not Be Copied Blindly

This is not a universal recipe. With partial offloading of layers to the CPU, a large context, handling multiple requests or parallel tasks, limiting the processor can already reduce speed. The result also depends on the LM Studio version, model format, launch parameters, drivers and the specific cooling design.

Be especially careful when changing PROCHOT and fan control. Wrong thresholds can lead to overheating, instability or accelerated component wear. Someone else's configuration file cannot be considered safe for a different laptop model.

A workable approach is to first record the baseline figures: generation speed, CPU and GPU load, temperatures, power consumption and noise. Then disable only turbo boost, repeat the same request and compare the results. If tokens per second did not drop while the temperature fell noticeably, you can carefully tune the fan curve. The main takeaway of the experiment is not the specific numbers but the method: limit the resource that consumes power but does not speed up your actual workload.