RTX 4090: Open-source trick lifts AI throttle, up to 39% more speed

What this is about

Community hobbyists have found a way to remove a hidden throttle in Nvidia's consumer graphics cards. According to Netzwelt (citing Geeky Gadgets), this can boost the AI performance of an RTX 4090 by up to 39 percent. It relies on a free Linux kernel module, not supported by Nvidia, that bypasses restrictions in the firmware. This isn't about higher clock speeds. Instead, certain compute units don't run AI workloads at the speed the hardware is apparently capable of.

The hack is unofficial and not without risk. If you want to use it, you need technical know-how, you lose your warranty and you have to switch off security features.

What exactly is throttled

Based on a video from the channel Kai, Geeky Gadgets describes several reasons why real-world AI performance falls short of the advertised figures. For the RTX 4090, Nvidia quotes around 1,321 AI TOPS (trillions of operations per second). That number rests on assumptions:

• Sparsity: The values apply to "sparse" models in which parts of the computation are skipped. That doesn't hold for every real-world workload.
• Preference for low precision: The hardware is optimized for formats like INT8, while full-precision (FP32) calculations are put at a disadvantage.
• Register throttling: Firmware restrictions slow down 32-bit accumulation, i.e. summing intermediate results at high precision.

Tom's Hardware describes the same mechanism as a "half-speed limiter". Two registers that select the speed of the tensor units are protected by a write lock. The unlocker gets Nvidia's signed booter firmware to open that lock, so the tensor units run at full rate. Readers in the comments also suspect that Nvidia reserves this performance for professional cards, but Nvidia has offered no evidence of this. Netzwelt sees pricing strategy as a possible reason and points to Tech Times, which says such firmware differences between professional and consumer products go back to the 2000s.

The Linux module and its effect

According to Geeky Gadgets, the module exploits a firmware vulnerability. It reports 33 to 39 percent more speed on full-precision workloads, such as:

• High-resolution image generation
• Vision-language models and long-context language models
• Fine-tuning AI models

Energy efficiency is said to improve because the card gets closer to its advertised performance. There is little benefit for quantized models (such as GGUF in llama.cpp, where numbers are stored more coarsely), memory-bound tasks like token generation, gaming and crypto mining. According to Netzwelt, anyone hoping for better gaming performance will be disappointed.

The new Windows-friendly variant

Tom's Hardware presents the project "nv-uefi-unlocker", which makes the trick much easier. It boots from a FAT32-formatted USB stick before the Nvidia driver loads, and works on both Windows and Linux. No VBIOS flash or patched driver is needed. It uses the community payload V67 with readback verification, which is meant to make the early experimental versions safer.

The package contains no Nvidia firmware. Instead, a script downloads the public Nvidia driver packages (version 610.43.03), verifies them and rebuilds the firmware files locally. The project builds on Zagitta's "gpu-unlock", the RTX 40 fork "gpu-unlock-ada" and an earlier Windows EFI tool. The process: download the release, run the script, copy the files including BOOTX64.EFI to the stick, boot from it via the one-time boot menu and confirm with the space bar within 30 seconds. After that, the normal bootloader takes over.

The catch: The unlock doesn't survive a power loss, so you have to repeat it on every boot. It requires UEFI mode and Secure Boot either disabled or set to trust a self-signed EFI file. According to one developer, game anti-cheat that requires Secure Boot can't be worked around. RTX 20 owners must first run a script that switches the Windows driver into GSP firmware mode. On laptops, dGPU-only or MUX mode is recommended. On Linux, power management can shut the GPU down when idle and silently erase the unlock.

Measurements and compatibility

Tom's Hardware lists results (FP16/BF16 with FP32 accumulation). According to the article, they aren't comparable between cards because the tests vary:

Some of these figures are well above the 39 percent from the Linux report, but they depend on the test. On compatibility: RTX 20 is "fully compatible, some models with known issues" (the RTX 2080, 2080 Ti and 2080 Super fail at boot or at handoff). RTX 30 is fully compatible, and RTX 40 is fully compatible on desktop and laptop. GTX 16 is theoretically compatible but unverified. Pre-Turing cards, GA100 and the RTX 50 series are not supported. According to Tom's Hardware, the team is struggling with "far more security" than expected there.

Risks and alternatives

• Security: Secure Boot has to stay off, which makes the system more vulnerable to attacks.
• Warranty: It is voided, and Nvidia doesn't cover damage.
• Fragility: Nvidia could close the loophole with a firmware update. According to Netzwelt, the company hasn't responded so far.
• Performance limits: Results depend on power and temperature limits. According to the project, combining it with the "nvoc" power-limit unlock lets you exceed factory limits. A per-card ROM hash check, which the project's safety rules require, hasn't been implemented yet.

As a simpler alternative, Geeky Gadgets suggests raising the GPU's power limit, for example from 200 W to 250 W or more. That can add speed on power-hungry workloads without touching the firmware.

Our take

The find reveals a gap between advertised and practical performance on consumer cards that matter for local AI. It's mainly worth it for Linux or Windows users with full-precision workloads, and only if they can bear the risks. A caveat: The 39 percent figure comes from a video report, and the Windows numbers are mostly owners' own measurements. Nvidia's motive for the throttle is unproven (speculation about product segmentation), and there is no official statement. The price of at least €4,999 at idealo cited by Netzwelt is also just a snapshot. The ComputerBase report on a regular Nvidia firmware update against boot problems on the RTX 4090/4080 is about a different issue and has nothing to do with the unlock.