this is not a scam. the way it works is: when nvidia makes their chips, they just make enterprise level chips because it’s cheaper to manufacture that way. and they use software to make the consumer edition GPU less stronger. all this mod is unlocks that software and run it like enterprise level GPU.
it only works with 30xx cards but has anybody tried this? and does it improve topaz performance?
I think most users here already know this, but I wanted to warn you guys against this trick:
It doesn’t bypass the board’s power limit, but it can push power consumption dramatically higher. Strictly for advanced users with top-notch cooling (STRIX-class or watercooled GPU).
Unlocking full FP16/BF16/TF32 Tensor throughput can significantly increase local power density within each Tensor Core cluster. These units occupy tiny areas within their respective SMs, potentially creating nasty thermal gradients. Monitoring the Hot Spot, not just the core temperature, is absolutely essential.
With pumped-out thermal paste or a mediocre heatsink, you’re gambling with some very expensive silicon.
I love pumping amperes into my GPUs, CPUs and even RAM, but this particular trick genuinely scares me.
And spare me the “the developer says it’s safe” argument. This hack is barely two weeks old. By the time people start screaming about fried GPUs, those reassurances will come a little too late.
My suspicion is that NVIDIA may reserve the dies with the best electrical and thermal headroom for its professional RTX cards, particularly for sustained matrix-heavy workloads.
Such practices have been extensively documented in the past across virtually all chip manufacturers.
Just look at the AD102 binning for example:
Note: I mentioned Ada because this trick already works on Turing, Ampere, and Ada under Linux. The current Windows port only supports Ampere, but that’s just the first step.
Undervolting is your friend 
I don’t know anything about this but, I see Nvidia’s CEO Jensen Huang has been caught Playing Games 
Higher heat density in the Tensor Cores regio? That could be, but then power consumption in other areas of the chip would have to be lower, since the maximum TDP is fixed.
I’m not sure NVIDIA actually uses better-binned chips for its professional cards. For example the 5090 and PRO 5500 Blackwell both use the same GB202 chip and have the same 600W TDP. However, the PRO 5500 comes with 84 GB of GDDR7 ECC memory, which consumes more power than the 5090-32GB. So the Pro5500 GPU itself must consume slightly less power to stay within the same TDP.
What I’m trying to say is that the 5090 is pushed at least as hard as the PRO 5500, so its chip shouldn’t necessarily be of lower quality. Normally, manufacturers bin their chips during production, with lower-quality silicon typically running at lower clock speeds.
Anyway, I don’t know for sure either. It’s just a thought experiment. What can certainly happen, though, is that the GPU hits its TDP limit more often with the hack than without it.
What is clear, that professional GPUs in servers often run with liquid cooling in air-conditioned server rooms at around 20°C, while ours have to deal with office temperatures of up to 35°C in summer. That’s why I think undervolting is definitely not a bad idea, especially when the GPU is running nonstop.
You’re confusing board-level power limits with local silicon power density. The unlock can increase total power consumption without exceeding the TDP. Even at identical wattage, where the heat is generated matters.
Also, the RTX 5090 is rated at 575 W, versus 600 W for the PRO 5500. More VRAM doesn’t automatically mean higher memory power consumption.
Binning involves voltage/frequency characteristics, leakage and stability margins, not just clock speeds.
Undervolting certainly reduces heat and can provide additional clock headroom. But it doesn’t change the silicon’s intrinsic electrical characteristics. A poorly binned die doesn’t magically become a golden sample.
After 40 years of overclocking, I’m well aware of the distinction. 
My point is that binning could play a role in NVIDIA’s Tensor Core restrictions. I never claimed it was proven.
Tensor Cores are utilized more heavily after being unlocked does not, by itself, prove that this actually creates dangerous localized temperature gradients.
Establishing that would require measurements or reliable thermal simulations. Furthermore, a standard GPU hotspot sensor cannot guarantee that every tiny, localized temperature spike is detected.
In the case of the RTX 5090, we now know that the same chip is also used in a professional graphics card, where it delivers full FP16/FP32 performance.
I think you’re basically technical right. The question is whether it will become significant enough to actually cause a problem. I think with undervolting, you should be in the clear anyway, since power consumption also decreases in the hotspot area.
Honestly, I’d be absolutely delighted if this hack turned out to be perfectly safe. I have nothing against it — quite the opposite! 
My point is simply that we don’t have enough evidence to establish its long-term safety yet. Undervolting certainly helps, but it doesn’t eliminate every potential risk.
We’ve seen enough AMD I/O dies fail after months, sometimes even years, effectively killing the entire CPU for reasons nobody initially suspected.
I’m not suggesting the same thing will happen here. I’d just rather let time and proper testing settle the question before gambling with expensive silicon.
Believe me, I genuinely hope this hack proves me overly cautious!
Note: To my knowledge, NVIDIA has never publicly disclosed the exact number or placement of its on-die thermal sensors.
The 9 and 12 readings identified through reverse engineering represent thermal registers, not necessarily individual physical sensors.
Some values may already be aggregated, averaged or otherwise processed.
Even NVIDIA’s public API doesn’t reveal the actual sensor topology.
Therefore, we simply don’t know how accurately localized Tensor Core hotspots are monitored.
nobody has answered the most important question here though. would this mod/unlocking even help with topaz performance? YES or NO
Try it and let us know.
The answer is absolutely yes. On my RTX 4070tis, the speed has been increased from 0.74fps to 1.0 fps. I also set a fixed 850mv@2520MHz and VRAM+1750 in MSI afterburner which controls the power at 230W. I don’t know the temperature for hot spot but general temperature is around 66°C in a 26°C room.
netlearner i thought this mod only works with 30xx. how did you get it working on your 40xx card ?
and also what model are you using that has been increased from 0.74 fps to 1 fps ?
All RTX 20-40 can get benefit from it except RTX 50s. I am using SLP 2.6, but with 3rd party launcher. If it is Topaz factory default, it should be 0.54fps.
i don’t understand: are you using GitHub - windows-GPU-UNLOCK/Windows-GPU-UNLOCK: Windows GPU UNLOCK: RTX3090 UEFI port, Korean no-reset Windows boot guide, verified source and benchmark evidence · GitHub this or something else?
Not this one, it’s NV-UEFI-Unlocker
netlearner. i see, thank you for this info. how does this unlocker/mod do on models like rhea/nyx/aion ?
I didn’t test those models. I guess any model utilize Tensor core will get benefit for sure.
are you telling you don’t use any other models except starlight which can only do 1 fps?
Yes, at this moment I only have one requirement to upscale an old K-drama and my current workflow is SLPX1 + TheiaX2. Some users say RheaX2 has better quality. I agree, but still don’t find a proper parameter to cleanup the monster face problem with RheaX2 although I already use SLP sharpness 1 to minimum the aggressive skin texture.
i see. thanks for the quick response. i also have issues with rhea creating monster faces. does starlight do a lot better with “not creating monster faces”? i don’t have proper hardware to use starlight
