Open-source tool GPU Unlock unlocks hidden compute power in RTX 30 series graphics cards

Open-source tool unlocks hidden AI compute: RTX 30-series older cards "revived" Source: Bandaozongheng Published: 2026-10-26 13:45 NVIDIA compute-generated poster An open-source tool called GPU Unlock can unlock the throughput limit of RTX 30-series graphics cards in Tensor Core matrix computation, allowing older graphics cards to release hidden AI compute power.

Over the past few years, NVIDIA has consistently drawn a clear line in software between consumer-grade graphics cards and professional-grade graphics cards. But recently, an open-source project has "erased" this line.

An RTX 3090 Ti released more than four years ago saw its BF16 compute power soar from 87.5 TFLOPS to 157.4 TFLOPS, an increase of nearly 80%; actual image generation task throughput increased by about 60%—this is not a new card launch, but rather the open-source project GPU Unlock "unlocking" the artificially limited Tensor Core instruction issue rate on consumer-grade Ampere graphics cards. No VBIOS flashing and no driver modification are required; it can be achieved with just a Linux kernel module that rewrites registers before the driver loads.

Ampere compute power that was "soft-throttled"

GPU Unlock was developed by GitHub user Zagitta and targets an instruction issue rate limit that NVIDIA set on consumer-grade Ampere graphics cards: in FP16/BF16 Tensor Core operations using FP32 accumulation, the issue rate of GeForce graphics cards is significantly lower than that of professional compute cards with the same architecture.

In terms of technical approach, the project combines previous firmware unlocking technology for CMP mining graphics cards with an early initialization scheme. Through a Linux kernel module, before the NVIDIA driver binds to the graphics card, it writes to the speed select registers (SS0=`0x88888888` selects full BF16 rate, SS1=`0x8` selects IMLA4), performs a function-level reset (FLR), and then the original driver takes over the graphics card normally.

This means there is no need to re-patch every time the driver is updated. Currently, this project supports GA102, GA103, GA104, and GA106 cores, covering the vast majority of RTX 30-series models.

After "unlocking," how much faster can it get?

According to the project's GitHub repository and a large number of actual tests submitted by community users, the compute power gap before and after unlocking can be described as the difference between "crippled" and "full-blooded."

The flagship model RTX 3090 Ti, under a 450W power limit, saw BF16 matrix operation performance jump from 87.50 TFLOPS to 157.41 TFLOPS, an increase of about 80%; TF32 performance rose from 42.58 TFLOPS to 82.61 TFLOPS, nearly doubling (+94%). Even the lower-power RTX 3080 10GB, while maintaining the factory 320W power limit, saw its FP16/BF16 throughput increase from about 65 TFLOPS to the level of 110 to 113 TFLOPS.

The general feedback from community users is that on RTX 30-series cards, BF16 performance improvements generally range from 70% to 80%, and the RTX 3090 in PyTorch actual tests can also jump from 75 TFLOPS to 135 TFLOPS, an increase of about 67%.

Of course, the benefits of unlocking have also spread to other architectures. The FP16/BF16 instruction throughput performance of the RTX 4090 directly doubled from 165.6 TFLOPS to 331.6 TFLOPS, and the FP16 performance of the RTX 2080 Ti also increased from 59.9 TFLOPS to 111 TFLOPS, an increase of about 1.9 times.

In local large language model (LLM) inference scenarios, some RTX 3090 users measured under vLLM workloads that prompt prefill throughput increased by about 33%, and first-token latency decreased by about 25%. However, it should be noted that the subsequent Decode stage of token-by-token generation shows almost no obvious improvement, because that stage is more limited by memory bandwidth rather than Tensor Core compute power.

In image generation, the benefits are more intuitive. When RTX 3090 Ti users run BF16-precision models with ComfyUI, throughput improvements can reach 56% to 59%; among them, MOSS-Image SFT BF16 improved by 58.6%, and Qwen-Image 2.1 Turbo BF16 improved by 57.2%. This almost means that on the same hardware, the time to generate an image can be shortened by nearly 40%.

Power consumption, risks, and spillover effects

A doubled instruction issue rate means higher power consumption. Only a 450W 3090 Ti can reach 157 BF16 TFLOPS; if the power wall is pressed down to 225W, the BF16 improvement of the 3090 is only about 25%; the RTX 3060 at 170W shows only about a 6% improvement in nanochat training. Whether the cooler, power supply, and power connector can keep up is a question users must evaluate themselves. Regarding long-term reliability and warranty, this operation is a firmware-level modification not officially supported by NVIDIA, and risks are borne by the user.

It is worth noting that similar technology has already spilled over to the Ada architecture. The independent project GPU Unlock Ada achieved FP16 on the RTX 4090 from 165.6→331.6 TFLOPS, nearly doubling; the RTX 4070 Ti SUPER also has experimental success cases. On Windows, there is currently only an experimental implementation based on UEFI, and it is not yet universal.

Summary

In short, the essence of GPU Unlock is to restore the Tensor Cores of consumer-grade Ampere graphics cards, which were "soft-throttled" at the factory, to the full speed state that the hardware should have had. For users holding RTX 3090/3090 Ti for local AI inference, image generation, or model training, this is equivalent to getting an AI workstation with "nearly doubled compute power" at zero cost. But it does not benefit from memory bandwidth dividends, provides no gain in gaming, and requires bearing power consumption, cooling, and warranty risks. Against the backdrop of continuously spilling-over AI compute demand and the rising popularity of the second-hand graphics card market, how many older cards this "open-source key" can give a second spring to still needs longer-term verification by the community.

This article is reproduced from media reports or online platforms and represents the author's personal stance or views. Our reproduction is solely for sharing and does not mean that we agree with or endorse it. If the source attribution is incorrect or your lawful rights and interests are infringed, please contact customer service promptly, and as a neutral platform service provider we will promptly correct, delete, or handle it in accordance with the law. Comments No user comments yet Latest News More Related Recommendations

DB HiTek IATF 16949 recertification, focusing on automotive-grade chips 2026-10-26 13:54

Older cards are getting more expensive to rent! NVIDIA H200 rental price surges 143%: A100 rent also rises accordingly 2026-10-26 13:50

Samsung's self-tested Exynos 2700 performance is about 20% higher than Qualcomm's flagship chip 2026-10-26 11:48