3 mins
If you own a GeForce RTX 3090 or RTX 4090 and run AI models at home, you may have seen claims this weekend that modders unlocked twice the performance. The claim holds up, but only for a narrow kind of work. Open-source tools now lift a limit NVIDIA places on certain Tensor Core math on RTX 20, RTX 30 and RTX 40 cards. VideoCardz and TechPowerUp both reported the results on October 25, based on the benchmarks the projects published. Neither outlet reported a response from NVIDIA.
What is the RTX tensor unlock?
TechPowerUp explains that GeForce RTX 30 and RTX 40 cards use the same chips as NVIDIA's professional GPUs, but the GeForce versions issue certain Tensor Core instructions at a lower rate. The unlocks change low-level GPU configuration registers before the NVIDIA driver loads, so the card issues those instructions at the rate of its professional counterpart.
According to VideoCardz, nothing changes on the hardware and no VBIOS is flashed. The change is temporary: it's cleared as soon as the GPU loses power.
Who is behind it?
The work builds on earlier attempts to unlock NVIDIA's CMP mining GPUs. Developer Zagitta adapted those techniques for RTX 30 cards in the open-source GPU-Unlock project, and work involving TOTORONG's firmware payload extended support to RTX 40 (Ada Lovelace). A separate unlock for RTX 20 (Turing) followed the same approach.
How much faster is it, really?
In synthetic compute tests, the gains are large:
- RTX 3090 Ti: BF16 matrix multiplication went from 87.5 to 157.4 TFLOPS.
- RTX 4090: FP16 Tensor Core instruction throughput doubled, from 165.6 to 331.6 TFLOPS. TechPowerUp notes that this level is normally reserved for professional cards like the RTX 6000 Ada.
- RTX 2080 Ti: KohakuBlueleaf showed 111 TFLOPS, up from 59.9 TFLOPS. That card had already been physically modified to 22GB of VRAM, which has nothing to do with the software unlock.
Real apps gain less. VideoCardz says RTX 3090 owners report roughly 14% to 36% faster prompt processing in large language models, while token generation barely changes because it's usually limited by memory bandwidth. Some image-generation workloads can also speed up. TechPowerUp adds that board power stayed at the same 450 W before and after the unlock on the GA102 and AD102 chips.
Does the unlock boost gaming FPS?
No. VideoCardz is clear that gaming performance is unaffected. The unlocks mainly target FP16 and BF16 operations with FP32 accumulation, plus some TF32 and FP8 workloads, which is the math behind AI models, not game rendering.
Which GeForce cards are supported?
TechPowerUp lists the GA102, GA103, GA104 and GA106 chips for the RTX 30 unlock, and AD102, AD103 and AD106 as officially unlocked for RTX 40. VideoCardz reports successful tests on the RTX 3060, RTX 3070, RTX 3080 and selected RTX 40 models, but support isn't validated on every variant yet. The RTX 50 series (Blackwell) has no confirmed working unlock so far.
Is it safe to try?
It's experimental. Most projects target Linux, with experimental Windows versions, and VideoCardz warns that unsupported configurations may fail to boot. TechPowerUp adds that any GPU modification voids the warranty, and the unlock has to be applied at boot, before the NVIDIA driver takes control of the card. If your PC is mainly for gaming, there's no reason to take that risk.
Will used RTX 3090 and RTX 4090 prices go up?
Nobody knows yet, and there's no price data on it. VideoCardz points out that the RTX 3090 and RTX 4090 already attract AI buyers because of their 24GB of memory, and that faster compute could make them even more attractive on the used market. Some users are also asking NVIDIA to remove these limits officially on older cards that are no longer made.
Should you buy a GPU now or wait?
For gaming, this news changes nothing. Pick a card for its gaming performance and VRAM, and don't pay extra for a used RTX 3090 or RTX 4090 because of an unlock that doesn't add a single frame. If you already own one of these cards for local AI, the gains in prompt processing are real, but weigh them against the warranty and stability risks before you try.
Shopping for an upgrade? Browse our graphics cards, check our ready-to-play gaming PCs, or save with one of our refurbished PCs.
FAQ
Does the RTX tensor unlock increase gaming performance?
No. It speeds up certain AI math on the Tensor Cores. Reported gaming performance is unchanged.
Do I need to flash my VBIOS?
No. The tools change GPU settings before the driver loads, and the change disappears when the card loses power.
Does it work on RTX 50 series cards?
Not yet. There's no confirmed working unlock for Blackwell GPUs so far.
Will it void my warranty?
TechPowerUp says any GPU modification voids the warranty, so assume it does.
