Friday, August 28th 2026

DLSS 5 Patched to Work on RTX 4000 "Ada Lovelace" GPUs Despite No Official Support

The modding community has been hard at work ever since the untimely release of DLSS-NR, a precursor to DLSS 5. Now, modders have gotten NVIDIA's leaked DLSS 5 tool working on RTX 4000 series cards, even though the tool was only built for RTX 5000 Blackwell hardware. The DLL leaked earlier this week via an early-access build of NBA 2K27, and people on the RenoDX Discord quickly got it running in other games using ReShade and RenoDX.

As we noted in our DLSS-NR breakdown, the AI model runs on FP8, which Ada Lovelace GPUs support just fine, but the compiled binaries inside the file only targeted Blackwell's sm_120 architecture, with nothing built for Ampere or Ada. A modder going by "Uncle Burrito" has apparently fixed that by hunting down the incompatible binaries and swapping in ones built for Ada. Tom's Hardware confirmed the patch works on an RTX 4080, and said other developers in the same Discord are working on their own versions of the fix.
It's still early and rough around the edges, and there's no word yet on whether NVIDIA plans to officially support older cards at launch. However, given how quickly modders cracked this open, it looks like NVIDIA may have little choice but to consider it. Older cards like Ampere could be a tougher case though, since they lack native FP8 support altogether.
Sources: Uncle Burrito on YouTube, Tom's Hardware
Add your own comment

33 Comments on DLSS 5 Patched to Work on RTX 4000 "Ada Lovelace" GPUs Despite No Official Support

#1
sam_86314
If it's just running a small FP8 diffusion model, maybe modding in RDNA4 support is on the horizon?
Posted on Reply
#2
son_of_wasps
sam_86314If it's just running a small FP8 diffusion model, maybe modding in RDNA4 support is on the horizon?
Probably not because it's compiled for CUDA and tensor cores.

Theres a compatibility layer out there for CUDA but those are always finicky and it may not be optimized enough for real-time usage. www.xda-developers.com/nvidia-cuda-amd-zluda/
Posted on Reply
#4
Fatloader
Lucifer Mornignstar30% performance hit on rtx 4070ti.
Really? That's a lot lower than what me and a lot of others were expecting.

Would love to see performance numbers if they're available.
Posted on Reply
#5
son_of_wasps
FatloaderReally? That's a lot lower than what me and a lot of others were expecting.

Would love to see performance numbers if they're available.
I'm wondering if, unlike traditional rendering overhead which proportionally decreases as gpu powe goes up, DLSS 5 has a flat percentage overhead.

If a 5090 can do a base 60fps or smth, then the AI model also has to run at 60 FPS, therefore taking up more performance. If a 4070ti can only do 30 fps, then the model also only has to run at 30 fps. It's the same overall ratio of the performance, because DLSS 5 processing cost (and prior DLSS for that matter) depends on the power of the underlying graphics card.
Posted on Reply
#6
Lucifer Mornignstar
FatloaderReally? That's a lot lower than what me and a lot of others were expecting.

Would love to see performance numbers if they're available.
Honestly was hoping for half of what it is right now, but the good thing no matter the sliders seems to stay at pretty much locked 70fps in Stellar Blade vs 100fps before in same area with reshade rtx preset enabled as well. Moving sliders around doesn't seem to influence performance, but was just a quick test to see if it works at all, will check it out tomorrow properly. So far 30% seems fine for the quality improvement it offers.
Posted on Reply
#7
sam_86314
son_of_waspsProbably not because it's compiled for CUDA and tensor cores.

Theres a compatibility layer out there for CUDA but those are always finicky and it may not be optimized enough for real-time usage. www.xda-developers.com/nvidia-cuda-amd-zluda/
I messed with ZLUDA quite a bit when trying to get SDNext working in Windows on my 6800 XT; I recall it worked pretty well.

Apparently, they rebooted development on it and a lot has changed since I really messed with it.
Posted on Reply
#8
Legacy-ZA
I think a lot of people just don't know how to configure it correctly with the sliders, this is just me dabbling quickly, I suppose I could add a little more colour, perhaps they just want to make the tech look bad? The only problem I see is that, you guessed it, you hit the 16 GB VRAM limit, and lose way more performance than I should.

Look here:






*Edit*

There we go, perhaps a little close to the original? :)
Anyways, fun. try it folks :D




Some more:








So far, all I can say is, if you don't have a 5090 with that extra VRAM, you won't use this, you can argue you can turn on Frame Generation, but it will just be even more load on the memory buffer. !@#$ing nGreedia, I always said, that the 5070Ti / 80 should have had 24GB, this just proves it once again. I hate them SSSOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO damn much.
Posted on Reply
#9
oxrufiioxo
Legacy-ZAI think a lot of people just don't know how to configure it correctly with the sliders, this is just me dabbling quickly, I suppose I could add a little more colour, perhaps they just want to make the tech look bad? The only problem I see is that, you guessed it, you hit the 16 GB VRAM limit, and lose way more performance than I should.

Look here:






*Edit*

There we go, perhaps a little close to the original? :)
Anyways, fun. try it folks :D




Some more:








So far, all I can say is, if you don't have a 5090 with that extra VRAM, you won't use this, you can argue you can turn on Frame Generation, but it will just be even more load on the memory buffer. !@#$ing nGreedia, I always said, that the 5070Ti / 80 should have had 24GB, this just proves it once again. I hate them SSSOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO damn much.
My guess is the finished version will use Neural texture compression to make it more viable on 16GB cards. Only blackwell supports it so my guess is this will stay 50 series exclusive with mods for 40 series.

Probably why it's taking so long for Nvidia to give an official launch date.

Still taking a 30-50% hit to performance is a na dog from me unless the official version is substantially better than the screens/videos I've seen posted.
Posted on Reply
#10
Dragam1337
Legacy-ZAI think a lot of people just don't know how to configure it correctly with the sliders, this is just me dabbling quickly, I suppose I could add a little more colour, perhaps they just want to make the tech look bad? The only problem I see is that, you guessed it, you hit the 16 GB VRAM limit, and lose way more performance than I should.

Look here:






*Edit*

There we go, perhaps a little close to the original? :)
Anyways, fun. try it folks :D




Some more:








So far, all I can say is, if you don't have a 5090 with that extra VRAM, you won't use this, you can argue you can turn on Frame Generation, but it will just be even more load on the memory buffer. !@#$ing nGreedia, I always said, that the 5070Ti / 80 should have had 24GB, this just proves it once again. I hate them SSSOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO damn much.
I do have a 5090, and there is no way im gonna use this ai slop filter...
Posted on Reply
#11
john_
This looks like the case with FSR4 unofficially working on RDNA2. Let's see if Nvidia will offer it to RTX 40 owners officially.
Posted on Reply
#12
Red Hood
john_This looks like the case with FSR4 unofficially working on RDNA2. Let's see if Nvidia will offer it to RTX 40 owners officially.
nvidia stated that Ada does support neural rendering and as this dlss 5 model runs on FP8 which it supports



Posted on Reply
#13
oxrufiioxo
Red Hoodnvidia stated that Ada does support neural rendering and as this dlss 5 model runs on FP8 which it supports



I think the main issue on 40 series will be vram, 50 series can use Neural texture compression.

Will the 24GB on the 4090 be enough, maybe but 16 is likely gonna be a stutter fest.

Still a feature that basically halves performance is going to be DOA not even framegen will be able to save it in any semi demanding game. Hopefully the official release is substantially better for those excited about this tech.
Posted on Reply
#14
phanbuey
oxrufiioxoI think the main issue on 40 series will be vram, 50 series can use Neural texture compression.

Will the 24GB on the 4090 be enough, maybe but 16 is likely gonna be a stutter fest.

Still a feature that basically halves performance is going to be DOA not even framegen will be able to save it in any semi demanding game. Hopefully the official release is substantially better for those excited about this tech.
IMO main issue is they're going to have a hard time selling 6xxx series cards at $15,000 a pop so they need every reason possible to get people off their 5080s/4080s/4090s.

I guarantee the performance hit on 6000 series will be like 10%.
Posted on Reply
#15
Bigshrimp
Not that I actually care, but I doubt DLSS 5 will be officially supported by Nvidia with the RTX 4000 series. They will lock it in the 5000 series and beyond, like they typically do with any new tech they release.
Posted on Reply
#16
oxrufiioxo
phanbueyIMO main issue is they're going to have a hard time selling 6xxx series cards at $15,000 a pop so they need every reason possible to get people off their 5080s/4080s/4090s.

I guarantee the performance hit on 6000 series will be like 10%.
10% across the stack at general performance but very little hit with DLSS5 lol....

6090 might be ok but 3-4k MSRP... 6K+ street pricing with Fanboys saying Nvidia is da best 4eva.

We likely won't be seeing anything till well into 2028 though from Nvidia is my guess.
Posted on Reply
#17
wolf
Better Than Native
Legacy-ZAperhaps they just want to make the tech look bad?
There is 100% at least some of this going around.

Or just all it AI slop, then call it a day.
Posted on Reply
#18
Let3r
sam_86314If it's just running a small FP8 diffusion model, maybe modding in RDNA4 support is on the horizon?
Noooo. Please keep the slop away from my 9070 XT!
Posted on Reply
#19
john_
oxrufiioxoI think the main issue on 40 series will be vram, 50 series can use Neural texture compression.

Will the 24GB on the 4090 be enough, maybe but 16 is likely gonna be a stutter fest.

Still a feature that basically halves performance is going to be DOA not even framegen will be able to save it in any semi demanding game. Hopefully the official release is substantially better for those excited about this tech.
RTX 3090 is probably pretty fast to do FG in some very good way, but Nvidia chose not to give FG to such an expensive and fast card. The same could probably happen here. It isn't a VRAM, or Neural texture compression thing. It's just a corporate decision. Let's see if Nvidia says no, are we going to see Internet on fire, like with FSR4 on RDNA2/3? I bet my house on "NO".

And obviously Raytracing proved that "a feature that basically halves performance is NOT going to be DOA" even WITHOUT framegen. Today we have 6x FG. That's probably more than enough to push some gamers to buy an RTX 6080 or RTX 6090 at whatever the price.
Posted on Reply
#20
Hecate91
If this runs on FP8 there isn't any reason why Nvidia shouldn't add it to RTX 40 series cards, other than wanting to sell 50 series cards.
Otherwise this is indeed like FSR4 support on RDNA2 & 3, but I don't expect we'll see any similar complaints over 40 series not getting official support.
Posted on Reply
#21
oxrufiioxo
john_RTX 3090 is probably pretty fast to do FG in some very good way, but Nvidia chose not to give FG to such an expensive and fast card. The same could probably happen here. It isn't a VRAM, or Neural texture compression thing. It's just a corporate decision. Let's see if Nvidia says no, are we going to see Internet on fire, like with FSR4 on RDNA2/3? I bet my house on "NO".

And obviously Raytracing proved that "a feature that basically halves performance is NOT going to be DOA" even WITHOUT framegen. Today we have 6x FG. That's probably more than enough to push some gamers to buy an RTX 6080 or RTX 6090 at whatever the price.
There is a funny coincidence about both The int8 version of FSR 4 leaked well before AMD officially supporting it and made them look a lot worse because it worked mostly fine on RDNA3.

I have the same mentality with all features if the hardware supports them in a good enough capacity they should release it. I don't think 20 series should have gotten DLSS 4.5 it suck ass on them and should not be used.

I really only care about DLSS SR the one technology Nvidia makes that actually improves latency and gives you a meaningful performance boost.

The key difference is RDNA2/3 did not have a decent upscaler so blocking FSR4 was a pretty dick move this is gonna be mostly hated by gamers is my guess. Even if they think it looks ok does it look good enough to halve performance.

I wonder how Nvidia feels about 2K right now lol, I'm sure they are not very happy.
Posted on Reply
#22
TheDeeGee
And still i'm not going to use this tiktok slop filter.
Posted on Reply
#23
Lucifer Mornignstar
after trying it for more than 15 min, after some timem like 30 min it starts to consume more and more vram, untill the point of full memory and its not flushing it. sometimes happens sooner sometimes later. If you start to move sliders around a lot, it will just at one point cosume your whole vram and not let go. At which point have to restart the game to get performance back, but eventually will tank again, can see why this is no a full release yet.
Posted on Reply
#24
john_
oxrufiioxoThere is a funny coincidence about both The int8 version of FSR 4 leaked well before AMD officially supporting it and made them look a lot worse because it worked mostly fine on RDNA3.

I have the same mentality with all features if the hardware supports them in a good enough capacity they should release it. I don't think 20 series should have gotten DLSS 4.5 it suck ass on them and should not be used.

I really only care about DLSS SR the one technology Nvidia makes that actually improves latency and gives you a meaningful performance boost.

The key difference is RDNA2/3 did not have a decent upscaler so blocking FSR4 was a pretty dick move this is gonna be mostly hated by gamers is my guess. Even if they think it looks ok does it look good enough to halve performance.

I wonder how Nvidia feels about 2K right now lol, I'm sure they are not very happy.
The first time in my life that I noticed that artificial restriction from a company only to sell a different *newer* model, was with my ancient Epson Stylus Color 600 printer. Back they I had discovered that using the driver of 660 instead of 600, I was getting much better print quality with more controls over the colors the printer could produce. At the same period Nvidia was coming out with a driver promising 40% improvement in games and was commenting about that by praising Nvidia. Of course this is probably a 25+/- years old story, but it was showing how an establish company was doing business compared a company still fighting to grab the top.

I think hardware that costs hundreds or even thousands of dollars should be getting all the features it can support as long as it is still on sale. If I had an RTX 20 I would want DLSS 4.5 even if this feature is slow on that hardware, just to have the option to decide myself. You are right about FSR 4 offering a better quality option for RDNA2/3 owners and AMD should had released it long ago, but having played with FSR 3.1 on the RX 6600 and more specifically with AFMF that is upscaling and FG in the driver, not in game making it an even worst option, I can say that most will not really miss the quality of FSR4 because they will not know that they lose visual quality and frankly, if someone is enjoying the game, wouldn't care. There is huge exaggeration on tech sites about FSR 3 visual quality. Yes, if someone stops moving, cut speed at 50% and zoom 400% FSR 3 looks bad, but in gaming people wouldn't notice because we aren't talking about real images where the brain can spot and identify the artifacts or errors. A somewhat worst image in a game will look just like any image in any game.

Nvidia is checking public reactions before releasing what it will drive their future sales. This AI slop as most people call it, will be the killer feature that will be selling Nvidia graphics cards, or should I say, push people with even high end hardware to upgrade. This is going to be marketed and pushed by youtubers and tech channels as photorealistic graphics, as the mean to turn a computer game into a real life simulation and believe me many will swallow this easily and with a smile of pride on their faces. Nvidia feels fine with everybody who helps them to test public reactions before the official release. And seeing no one from the Nvidia fanbase demanding this to be officially released to RTX 40 owners, Nvidia can rest assured that they can ignore even RTX 4090 owners if they choose to do so, to almost no harm on them and their image.
Posted on Reply
#25
Raiden85
So Nvidia, what's your damn excuse this time? As expected, there's no hardware limitation; it's Nvidia being greedy and artificially locking it to the 5000 series.
Posted on Reply
Add your own comment
Nov 17th, 2026 13:51 CST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts