TL;DR
Thorsten Meyer AI published a tuning guide arguing that power limiting and undervolting should be the first heat-control step for high-power local inference workstations. The guide cites RTX 4090 measurements showing a 70% power limit cutting about 90 watts while keeping 93.4% of tokens-per-second performance.
Thorsten Meyer AI has published a workstation-tuning guide that recommends GPU power limiting and undervolting as the first step for reducing heat and fan noise in local AI inference systems, citing RTX 4090 workload data showing a large drop in power draw with a smaller loss in tokens per second.
The guide says local inference workloads often depend more on memory bandwidth than peak GPU core clocks. On that basis, it argues that cutting GPU power can lower heat output while preserving much of the throughput users care about: tokens per second.
In the cited RTX 4090 example, stock operation is listed at 390 watts, 72°C and 100% relative speed. A 70% power limit is listed at 300 watts, 67°C and 93.4% speed, while a 60% limit is listed at 260 watts, 62°C and 91.5% speed. The guide labels 70% as the recommended setting and 55% as the peak-efficiency point in its sample data.
The article distinguishes between simple power limiting and direct undervolting. It describes power limiting as the safer starting point because users restrict the card’s allowed power, while direct undervolting involves editing the voltage-frequency curve and testing stability under real workloads.
Why It Matters
The guidance matters for people running local LLMs because heat, noise and power draw are common limits in home and small-office AI workstations. If the cited pattern holds for a user’s hardware and model, a free software setting could delay or reduce the need for new cooling hardware, case changes or fan upgrades.
The performance framing is also specific to inference rather than gaming. The guide says gaming workloads can lose more performance when core clocks are reduced, while many inference workloads may remain closer to memory-bandwidth limits. That distinction is central to the claim that tokens per second can remain high after a power cut.
Thermal Grizzly WireView GPU – 1x8Pin PCIe Normal – GPU Power Consumption Measuring Device – PCIe Power Connector – Real Time Direct Monitoring – Made in Germany
REAL-TIME OLED WATTAGE: Instantly shows current GPU power draw in watts for quick, at-a-glance monitoring while gaming, benchmarking,…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background
The guide is presented as the first of five levers in a broader Thorsten Meyer AI series on reducing heat and noise in high-power AI workstations. It positions GPU power control as the highest-impact early step because it costs nothing and can be reversed.
The recommended workflow is simple: open a tuning tool, set a power limit near 70%, run the user’s real inference workload, measure power, temperature and tokens per second, then save the setting so it persists after reboot. The guide names MSI Afterburner for Windows and nvidia-smi or LACT for Linux setups.
“This is the first thing you should do to a high-power AI workstation.”
— Thorsten Meyer AI guide
“Local inference is memory-bound.”
— Thorsten Meyer AI guide
“Power limiting moves one slider and can’t damage anything.”
— Thorsten Meyer AI guide
VIPERA NVIDIA GeForce RTX 4090 Founders Edition Graphic Card
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What Remains Unclear
The results are not guaranteed for every system. The guide says figures vary by GPU, model, quantization, chip quality and workload. Stability under direct undervolting also remains workload-dependent; a curve that runs briefly may fail during a longer inference session.
The source material describes the RTX 4090 figures as measured from a sustained workload and refers to published RTX 5090 and RTX 4090 power-cap tests from 2025–2026, but it does not provide full raw test logs in the supplied material. Users still need to verify tokens per second, thermals and stability on their own machines.
Thermal Grizzly WireView Pro GPU – 1x12VHPWR Reversed – Advanced Power Meter for Graphics Cards – OLED Display – Temperature Sensors – Monitoring Tool – Made in Germany
Advanced power measurement device for graphics cards with 12VHPWR connector
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What’s Next
Readers applying the guidance would next test a conservative power limit, commonly around 70%, against their own inference workload and record the actual power draw, temperature and tokens-per-second result. Users seeking more performance per watt may then test a lower limit or a direct undervolt, with longer stability runs before saving the profile for daily use.
GPU Backplate Memory Radiator Aluminum Heatsink Cooler with Dual 70mm 4Pin PWM Cooling Fan and 1.2mm 7W Thermal Pad for Graphics Card RTX 3090
DUAL BALL PWM: 2x70mm dual ball bearing high performance gpu cooler fan (while others 50mm) for better cool…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is the main development here?
Thorsten Meyer AI has published a practical guide and infographic arguing that GPU power limiting and undervolting should be the first tuning step for reducing heat in local inference workstations.
Does the guide say tokens per second stay exactly the same?
No. The cited RTX 4090 example shows a small throughput loss at the recommended 70% power limit: 93.4% of stock speed while drawing 300 watts instead of 390 watts.
Is power limiting the same as undervolting?
No. Power limiting sets a maximum power draw and lets the GPU manage voltage and clocks. Undervolting directly changes the voltage-frequency curve and needs more careful testing.
Why might this work better for inference than gaming?
The guide says many local inference workloads are limited by memory bandwidth rather than core compute. If the GPU core is waiting on memory, reducing core power may have less effect on tokens per second than it would have on frame rates in a compute-heavy game.
What remains unclear for a specific user?
The exact result depends on the card, cooling setup, model, quantization, drivers and workload length. Users need to measure their own sustained run before treating any setting as stable.
Source: Thorsten Meyer AI
