Used 3090/4090 GPUs about to get more expensive again? An open-source unlocking tool pulls compute from 75 back up to 135 TFLOPS. There's new movement on the GPU price-increase front again. This time it's neither mining farm operators nor scalpers doing the work—it's a group of researchers. Someone found an unlocking method NVIDIA hid inside its GPUs, pulling back a big chunk of the locked-away compute on the RTX 3090, and power consumption didn't budge at all. The news

💡 What You Will Learn

Used 3090/4090 GPUs about to get more expensive again? An open-source unlocking tool pulls compute from 75 back up to 135 TFLOPS. There's new movement on the GPU price-increase front again. This time

Second-hand 3090/4090 GPUs Going Up Again? An Open-Source Unlock Tool Pulls Compute from 75 Back to 135 TFLOPS

There's new movement on the GPU price hike front again. This time it's not miners or scalpers making the move — it's a group of researchers. Someone found a way to unlock the compute NVIDIA hid inside its GPUs, pulling back a big chunk of the locked-away performance on the RTX 3090, all without touching power consumption. Now that word has spread, second-hand prices for the 3090 and 4090 will probably climb again.

Let's start with the most eye-catching numbers. Their tests ran on PyTorch, and on the same RTX 3090, with the power limit still at 360W, compute went from about 75 TFLOPS to 135 TFLOPS — roughly an 80% jump. On the 4090 side, a branch project pulled off something similar, with instruction throughput for the same kind of compute going from about 165 to 331 TFLOPS — nearly double. The results are impressive, and the approach is nothing like traditional overclocking.

What this method changes is a set of registers inside the GPU that control compute speed, called speed selectors. Power draw and clock speed stay untouched. These registers are normally locked behind permission protection and simply can't be written to at the system level. The researchers exploited a flaw in the GPU's boot firmware to grab write access before the official driver loads, set the speed tier to full-speed BF16, then did a reset to hand the card back to the stock driver. The whole thing is a Linux kernel module — no BIOS flashing, no driver modification.

One caveat needs stating up front: 135 TFLOPS is a paper peak — the ceiling after unlocking the rate tier for AI workloads like FP16/BF16 inputs with FP32 accumulation. No program will actually hit that number. When you really run large language models, what holds it back is usually memory bandwidth rather than tensor throughput, and community benchmarks typically show only about a 10% speedup; what genuinely benefits are workloads that lean on tensor speed, like graphics and training. In short, unlocking just gives you back the rate tiers the driver had locked away — it doesn't conjure up new cores out of thin air.

The most striking part is how it pokes a hole in Jensen's pricing logic. Gaming cards and professional cards use the same GA102 core, and per the whitepaper, the tensor compute rate for gaming cards is deliberately set at half of the professional cards. In other words, the "laser cut" is mostly just pricing the same core by tier — low tier for gaming cards, high tier for professional cards. This unlock hands that cut right back.

For a gamer with a 30-series card in hand, the barrier to entry is still there. Right now it's Linux-only, you have to compile the kernel module and deal with signing, and it has to be installed into the boot image before a cold start — a warm reboot won't work, and once the driver binds the card you can't touch it. The author himself notes it's only been validated on one or two RTX 3090 boards; it's fine as a benchmark toy, but expecting it to work across the whole lineup is premature.

The open-source project behind this is called gpu-unlock, hosted on GitHub (github.com/Zagitta/gpu-unlock), and strictly speaking it's still only a release candidate. The author combined the CMP unlocker technique with the timing of the UEFI approach; by default it doesn't touch any card — you specify which ones to unlock yourself via gpu_unlock.targets on the boot command line. It gained over two hundred stars within days of going public, and more people will follow. Before you actually hop on, it's worth reading through the limitations in the README so you know what you're getting into.