16 - Nov - 2026

Free firmware unlock nearly doubles AI speed on your RTX 30 and 40 GPU

Summary

  • Modders unlocked firmware limits, nearly doubling ML/AI tensor throughput on RTX 20/30/40 on Linux.
  • No gaming benefit; gains only affect FP16/BF16 AI math – helps local LLMs, upscalers, and ML workloads.
  • Linux-only, manual boot setup, and warranty void; could spike secondary market demand and prices.

Independent modders have released a series of firmware-level utilities that effectively circumvent the artificial limits NVIDIA places on its consumer cards. The result is a near doubling of the machine learning and AI compute performance of the RTX 20, 30, and 40 series hardware on Linux systems. You can try it out for yourself via the KohakUnlock repo, but it will void your warranty.

What this means for your GPU

It depends on how you use your PC

For the average user, it’s a long-standing pain point with corporate software gatekeeping. NVIDIA has traditionally built its consumer gaming graphics cards on the same silicon as its high-end workstation cards, but used firmware to limit consumer hardware’s raw mathematical throughput.

Starting with the GeForce RTX 30-series “Ampere,” a researcher created a tool that makes these graphics cards process certain math instructions twice as fast. This means cards like the GeForce RTX 3090 Ti can do matrix calculations up to 80% faster. This tool works on GeForce RTX cards with GA102, GA103, GA104, and GA106 chips.

ASUS TUF Gaming GeForce RTX 3090 graphics card box on a wood floor Credit: Amir Bohlooli / MUO

For the newer GeForce RTX 40-series “Ada Lovelace,” another tool doubles the speed for certain types of math used in AI. For example, with this unlock, the GeForce RTX 4090’s performance jumped from 165.6 to 331.6 trillion operations per second. Normally, only expensive professional cards like the NVIDIA RTX 6000 “Ada Lovelace” reach these speeds. It appears NVIDIA purposely slowed consumer cards for some AI tasks, but this unlock lets you restore their full power on AD102, AD103, and AD106 chips.

The newly developed open-pu-unlock and related modules intercept the boot process before official drivers bind to the kernel, overriding PCIe Function Level Reset restrictions. Rather than modifying the permanent video BIOS, the utility removes software bottlenecks on instruction-issuing rates for FP16 and BF16 matrix calculations. In benchmarks, high-end consumer cards like the GeForce RTX 4090 and RTX 3090 Ti achieve nearly double their previous tensor throughput, effectively matching workstation hardware that normally costs thousands of dollars more.

The engineering feat is certainly impressive, but the real-world impact on the average person is mixed and dependent on how you use your computer. PC gamers, in particular, should know these changes provide no benefit to traditional games, real-time rendering, or frame rates, since rendering engines already use the available graphics resources.

Instead, the main beneficiaries are hobbyists, developers, and researchers running their own local LLMs or doing the sort of heavy lifting you get with AI upscalers (e.g., AI upscaling of video), where the faster your maths throughput, the more tokens you can process.

At the same time, the community feels real unease about the secondary market. iameatingjam on the forum pointed out that if you’re going to unlock enterprise-level AI performance on an older consumer architecture, you’re going to create a lot of secondary market demand. If the hobbyists and small developers start to go out and pick up the older-gen consumer cards in order to build their own cheap ML rigs, then that’s going to further drive up the demand and the price and will effectively price out the average gamer who just wants to get an entry-level card. For the moment, they’re still Linux-only, require manual configuration at boot, and will void your hardware warranty.

Leave a Reply

Your email address will not be published. Required fields are marked *