GeForce Unlocked: Modders Boost RTX 20, RTX 30, and RTX 40 Compute Performance by Up to 2x - TechPowerUp

요약

NVIDIA's GeForce RTX 30-series and RTX 40-series graphics cards are based on the same dies as their professional-grade GPUs. This means the compute potential of consumer-oriented RTX GPUs is restricted by NVIDIA's firmware stack, but modders have managed to u…

본문

Sunday, October 25th 2026

GeForce Unlocked: Modders Boost RTX 20, RTX 30, and RTX 40 Compute Performance by Up to 2x by AleksandarK Sunday, 17:14 Discuss (44 Comments) NVIDIA's GeForce RTX 30-series and RTX 40-series graphics cards are based on the same dies as their professional-grade GPUs. This means the compute potential of consumer-oriented RTX GPUs is restricted by NVIDIA's firmware stack, but modders have managed to unlock the instruction issuing rate equivalent to their professional GPU counterparts. Starting with the GeForce RTX 30-series "Ampere," a researcher developed an unlock that effectively doubles the instruction issuing speed of FP/BF16 FMA into FP32 operations on GA102, GA103, GA104, and GA106 cards. As a result, cards like the GeForce RTX 3090 Ti experience up to 80% faster matrix multiplication in BF16, which is crucial for running nearly all neural networks. The unlock supports GA102, GA103, GA104, and GA106-based GeForce RTX graphics cards.

Next, we have the GeForce RTX 40-series "Ada Lovelace" with another project that has demonstrated a massive 2× speedup in FP16 and BF16 tensor throughput with FP32 accumulation. When the GeForce RTX 4090 was tested with FP16 tensor throughput and matrix accumulation, the GPU achieved 165.6 TeraFLOPS, while the unlock pushed that number to 331.6 TeraFLOPS. This is intriguing as these computational performance numbers are typically reserved for the higher-end professional segment like the NVIDIA RTX 6000 "Ada Lovelace." It seems NVIDIA artificially limited the AD102 silicon's performance in compute-sensitive matrix multiplication and accumulation, affecting the performance of running large language models. For this unlock, AD102, AD103, and AD106 are officially unlocked.

These two separate GPU unlocks appeared nearly simultaneously, with one demonstrated for the GeForce RTX 30-series "Ampere" and the other for the GeForce RTX 40-series "Ada Lovelace." Both unlocks successfully pass NVIDIA's driver checks and override PCIe Function Level Reset (FLR), running stock NVIDIA drivers, and CUDA benchmarks are performed normally. For example, the same 450 W TGP board power is used before and after the unlock on both GA102 and AD102. For those running compute-intensive tasks like local LLMs, the added compute power results in significantly higher token throughput in prefill and decode, allowing consumers to enjoy higher performance. However, it's important to note that any GPU modification will void the warranty and must be enabled at boot so these mods bind before NVIDIA drivers bind the card to the kernel.

Update 09:20 UTC: NVIDIA's GeForce RTX 20-series "Turing" are also being unlocked by the same technique, boosting the performance by 2x.

Sources: Ampere Unlock, Ada Lovelace Unlock, Turing Unlock

44 Comments on GeForce Unlocked: Modders Boost RTX 20, RTX 30, and RTX 40 Compute Performance by Up to 2x

#1 2080 ti too, btw. #2 Now we need this working in Windows #3

PurSpyk!!Now we need this working in Windows

won't help in gaming, only machine learning and AI stuff would benefit it and only on Linux so far has been tested, this would probably cripple prices of the 50 series over the 40 series. #4 In many ways Nvidia basically did nothing for the past 6 years except lock software features out and gimp their own products, what a remarkably greedy corporation.

#5 I think I’ll go with AMD in the future.
Between the whole AI situation getting out of hand and Nvidia’s 8-pin power cable design, I’ve had enough.

Raw performance is all I need now.
I’d rather not deal with "fake" FPS numbers anymore. #6 Everytime.

xxxZAURxxxRaw performance is all I need now.
I’d rather not deal with "fake" FPS numbers anymore.

Implying AMD is not doing the same and following Nvidia's footsteps with the so called 'fake' numbers and AI techs. #8 If its ends up making something like DLSS5 run better or some games, great. Otherwise, meh. #9 Can't wait until someone makes this compatible with Windows, would help massively with my LLM's :clap: #10 Sooo sick of news like this. So basically, there's now going to be people buying more 30 series gpus for ai workloads. fing great. That ought to help out the budget gamers get in the door. #11 The "Green" crowd might even end up thanking Jensen for this. #12 Free performance upgrade for the compute stack in the cards. Modders are doing amazing work, seriously. Almost every day modders are blowing up the news with all kinds of innovative techniques that gets around greedy companies restrictions. I am here for it. #13

RootinTootinPootinwon't help in gaming, only machine learning and AI stuff would benefit it and only on Linux so far has been tested, this would probably cripple prices of the 50 series over the 40 series.

Guess you forgot about dlss #14

SithaerImplying AMD is not doing the same and following Nvidia's footsteps with the so called 'fake' numbers and AI techs.

... and to still lose in raw performance anyways :) #15 This would be nice if ported over to Windows to work with my pair of memory-modded 2080 Tis I bought for dabbling in offline AI. #16 For now, it doesn't help ray-chasing games any only automatic idiots. #17 and if u read or even test ,they faked test and all numbers ...
wonder why ... cause on linux if u run any normal not faked test these cards crawl compared to nvidia numbers #18

SithaerImplying AMD is not doing the same and following Nvidia's footsteps with the so called 'fake' numbers and AI techs.

I have yet to see AMD do something like "5070=4090" BS yet. #19

Hecate91I have yet to see AMD do something like "5070=4090" BS yet.

Jensen is unbeatable when it comes to brainwashing tactics, lies, and scams.

Other corporations can only catch a glimpse, from the shadows, of this level of shady behavior. #20 That’s an enormous boost for ai workloads.

But I wonder how this affects power consumption and cooling/heat dissipation requirements. #21

lemonadesodaThat’s an enormous boost for ai workloads.

But I wonder how this affects power consumption and cooling/heat dissipation requirements.

"The gains are dense GEMM throughput from each report's own benchmark, so they are not directly comparable. The power limit matters a lot: the lowest figure (+25%) is from cards capped at 225 W."

They ran it with a TDP cap to avoid issues. #22

DenverJensen is unbeatable when it comes to brainwashing tactics, lies, and scams.

The leather-jacket man even sends his bots to comment "but no MFG, so this is irrelevant" on threads about AMD GPU driver updates.

Nvidia knows the meaning of the word "gatekeeping" very well. #23 With video cards becoming unobtainable, it's not much of a stretch to imagine the market will transform the way the car market transformed in Cuba, where you can see cars from the 60s on the streets today. And I'm only half-joking. #24

RootinTootinPootinwon't help in gaming, only machine learning and AI stuff would benefit it and only on Linux so far has been tested, this would probably cripple prices of the 50 series over the 40 series.

If it would help in applications like Topaz Video (AI based upscaling) it would be immensely helpful for me to perform this mod in Windows. Some of the models take several days of processing time. #25 Great news for Ai users with Nvidia GPUs on Linux. My current Linux build is AMD based, but this is very promising. Add your own comment

출처: Techpowerup.com

원문 게시일: 2026-10-25T08:14:32Z

이 글은 NewsAPI를 통해 자동으로 수집된 콘텐츠입니다. 원문은 위 링크를 통해 확인하세요.

GeForce Unlocked: Modders Boost RTX 20, RTX 30, and RTX 40 Compute Performance by Up to 2x - TechPowerUp

요약

본문

Related News

44 Comments on GeForce Unlocked: Modders Boost RTX 20, RTX 30, and RTX 40 Compute Performance by Up to 2x