It let's you change the NR resolution from 100% to 75%, 50% and 33%
But it drastically decreases the fidelity of the NR, imho it's already a heavy but usable impact at 75%, then becomes virtually useless below, even though it also runs way better.
I think the problem with the current implementations is that they can only run on the final frame
But the ideal way would be to run it together with something like DLSS, lets say render a 4k output at 1080p or so, apply the NR to that image, and then let DLSS do it's upscaling magic
That should dramatically improve performance while only marginally decreasing fidelity
Pretty sure that is how nvidia intended it, but obviously requires proper game implementation or a more in-depth game specific mod like PureDarks Skyrim upscaler
DLSS 5 launches on September 3 with NBA 2K27, gonna be interesting to see what the actual benchmark results are instead of how its implemented as a mod and with the pre-release version (though its of course possible nothing has changed).
Amadeus wrote:
What's that gonna run at, 5 fps and look weird and unstable af?
No way, its a 23 year old game so a 50% or more drop in FPS is nothing if you get like 400 FPS or whatever, which would pretty reasonable for such an old game as CoD. not sure what you mean with more passes, its with DLSS 5 on and off, not dual DLSS 5..
I have a hard time thinking the performance drop can be this much for actual DLSS 5 when its implemented for games. it would appeal to less than 1% of gamers (5090 users are 0.4% of Steam users), so its practically for no one, or its something for next generation of cards and to hype those up. We'll see.
Edit: talked a little to chapgpt and this is what it predicts in terms of performance
Most likely: 5070 Ti works quite well at 1440p, perhaps with a noticeable ~15–30% neural-rendering cost. DLSS SR + FG compensates for it. RTX 5090 gets the spectacular 4K showcase numbers.
Less pleasant: 5070 Ti works, but DLSS 5 is expensive enough that you really need aggressive upscaling + FG/MFG. NVIDIA can truthfully advertise support, while RTX 60 later gets much faster neural inference.
Worst case: current Blackwell is effectively a preview platform and DLSS 5 only becomes genuinely attractive on RTX 60. I think this is possible, but less likely.
DLSS 5 launches on September 3 with NBA 2K27, gonna be interesting to see what the actual benchmark results are instead of how its implemented as a mod and with the pre-release version (though its of course possible nothing has changed).
Amadeus wrote:
What's that gonna run at, 5 fps and look weird and unstable af?
No way, its a 23 year old game so a 50% or more drop in FPS is nothing if you get like 400 FPS or whatever, which would pretty reasonable for such an old game as CoD. not sure what you mean with more passes, its with DLSS 5 on and off, not dual DLSS 5..
I have a hard time thinking the performance drop can be this much for actual DLSS 5 when its implemented for games. it would appeal to less than 1% of gamers (5090 users are 0.4% of Steam users), so its practically for no one, or its something for next generation of cards and to hype those up. We'll see.
Edit: talked a little to chapgpt and this is what it predicts in terms of performance
Most likely: 5070 Ti works quite well at 1440p, perhaps with a noticeable ~15–30% neural-rendering cost. DLSS SR + FG compensates for it. RTX 5090 gets the spectacular 4K showcase numbers.
Less pleasant: 5070 Ti works, but DLSS 5 is expensive enough that you really need aggressive upscaling + FG/MFG. NVIDIA can truthfully advertise support, while RTX 60 later gets much faster neural inference.
Worst case: current Blackwell is effectively a preview platform and DLSS 5 only becomes genuinely attractive on RTX 60. I think this is possible, but less likely.
That 50% number is only a relative estimate for modern games
Of course the compute cost of something like DLSS 5 is many magnitudes larger than rendering a 23 year old game on a modern GPU, which will hit engine, driver and CPU limits first
So rather than going from 400 fps to 200 fps you're going to maybe 50-60 fps
There really is a limit on how many NR frames of a given resolution a RTX gpu can output
For example Deus Ex runs at like 600 fps but with NR only runs at around 70 fps, that's on a 5080
Because while the GPU can output many more rasterized frames than that, it can't produce more NR frames than that
PureDark managed to implement a way to select a different GPU for framegen and NR
So he let his 5090 run the raster on Skyrim
And a 4070ti running DLSS 5 / NR
It managed to do around 30fps
But I am sure official implementations will run better, esp. since they will likely run NR on frames before upscaling, so at a lower resolution
That Call of duty (2003) screenshot definitely ran multiple passes of DLSS 5 thanks to special mods that let you run it up to 10 times on each frame, which ofc will result in more extreme results, more artifacts and far less performance.
...
PureDark managed to implement a way to select a different GPU for framegen and NR
So he let his 5090 run the raster on Skyrim
And a 4070ti running DLSS 5 / NR
It managed to do around 30fps
...
Pretty cool though, that you can run it that way. But in this specific case wouldn't it be better to use the 5090 for the DLSS 5 part? Surely that would result in more FPS?
"Enlightenment is man's emergence from his self-imposed nonage. Nonage is the inability to use one's own understanding without another's guidance. This nonage is self-imposed if its cause lies not in lack of understanding but in indecision and lack of courage to use one's own mind without another's guidance. Dare to know! (Sapere aude.) "Have the courage to use your own understanding," is therefore the motto of the enlightenment."
(this is nit what DLSS5 normally looks like, its overblown.)
"Enlightenment is man's emergence from his self-imposed nonage. Nonage is the inability to use one's own understanding without another's guidance. This nonage is self-imposed if its cause lies not in lack of understanding but in indecision and lack of courage to use one's own mind without another's guidance. Dare to know! (Sapere aude.) "Have the courage to use your own understanding," is therefore the motto of the enlightenment."
"Enlightenment is man's emergence from his self-imposed nonage. Nonage is the inability to use one's own understanding without another's guidance. This nonage is self-imposed if its cause lies not in lack of understanding but in indecision and lack of courage to use one's own mind without another's guidance. Dare to know! (Sapere aude.) "Have the courage to use your own understanding," is therefore the motto of the enlightenment."
Actually lets you stack DLSS 5 up to 30 times and you can see why it has it's name
couleur wrote:
Amadeus wrote:
...
PureDark managed to implement a way to select a different GPU for framegen and NR
So he let his 5090 run the raster on Skyrim
And a 4070ti running DLSS 5 / NR
It managed to do around 30fps
...
Pretty cool though, that you can run it that way. But in this specific case wouldn't it be better to use the 5090 for the DLSS 5 part? Surely that would result in more FPS?
Yea PureDark did that only for testing, if you did it the other way around the 4070ti would not bottleneck the NR frames generated by the 5090 but of course at that point you could probably just use the 5090 for raster + NR and get about the same results
A better use-case would be to have two similarly powerful GPUs like a 5080 + 5070ti, 5090 + 5080 etc
or ideally two of the same GPU, then one can do a game all on raster and the other can spend its resources on framegen and NR
Of course Nvidia GPUs have dedicated on-chip hardware for the AI stuff so it might not make that much of a difference since that will be the limiting factor, also exactly why you might get 100-200 fps from raster and then NR pulls it down to 60-70 fps because that's all that the tensor cores and RT cores can generate.
Given that Nvidia managed to make this work on a single 5060 down from needing 2x 5090 the most likely performance upgrades will come from them optimizing the model.
So yeah.. 3070 died, stuck with old 2060 I had lying around.
I'm strongly leaning towards getting a 9070 or 9070 XT. Price/performance is a lot better than the performance equivalent ass-overpriced Nvidia stuff (5060 Ti and 5070 respectively).
Cheapest prices in Sweden right now (SEK converted to Euro):
Palit GeForce RTX 5060 Ti 16GB Infinity 3 - €840 (!!! )
PowerColor Radeon RX 9070 16GB Reaper - €700
-------------------------------------------------------------------------------------------
Acer Nitro Radeon RX 9070 XT OC 16GB - €750 (*second cheapest)
PNY GeForce RTX 5070 12GB OC - €840
(only 12GB on the 5070??? )
* - The reason I chose the second cheapest 9070 XT is because the cheapest (€680) is the Acer Predator BiFrost Radeon RX 9070 XT 16GB OC which is plagued with various problems: build quality issues, coil whine, crappy and noisy cooling solution leading to record high hot spot and VRAM temps which in turn leads to serious thermal throttling (~15-20% lower clocks due to power limitation).
How Acer could release two 9070 XT cards where one is like most 9070 XT cards while the other is a turd is beyond me. I guess the BiFrost was assembled in some remote child sweatshop in some Chinese village where they used the cheapest and crappiest materials they could get their hands on.
Anyway, if I bought a card today it would be one of the Radeon cards. They're still waaaay overpriced (VRAM costs and greed-flation).
Nvidia is just taking the piss on the consumers. Jensen (and his underling executives) is the greediest man in the computer industry and treats Nvidia like it's his personal game of Monopoly™ while he sucks investor cocks so he'll be "loved". If he crashed while flying over some ocean while taking a long detour due to some severe storms, got stuck on some remote and unmapped island where he'd have to eat durian fruit to survive, it would cause a massive drop in the valuation of Nvidia and most investors would panic and sell off a lot of their Nvidia shares. It would be almost impossible for Nvidia to replace Jensen with someone that has the same level of greed, personality and business skill.
There's currently nothing on the horizon that gives us any kind of hope of more sensible prices. Nvidia is running a very well constructed pricing scheme that visibly seems to be in a grey area as far as legality is concerned. He's probably been planning the current situation with the big RAM manufacturers and AI giants for years. Fuck him and the others. They set a precedent that infected the rest of the retail chain.
So GPU-1 renders the game frame and sends it to GPU-2 via PCIe, GPU-2 performs the DLSS5 processing and sends the resulting frame, again over PCIe, to GPU-1 where it's displayed on the screen.
That feels like at least a ~2 fps latency introduced. I don't know though, during my quick glance at the github text files I didn't see any clear mention about added latency, just that there is (obviously) latency in this method.
* note: technical rant incoming
I dropped a thorough description to the AI bot in Brave and got this:
Quote:
Added Latency: ~1.3–1.5 ms per frame
The dominant new cost is the PCIe 4.0 x16 round trip:
Component Time
Frame size (1440p RGBA) 2560 × 1440 × 4 ≈ 14.7 MB
PCIe 4.0 x16 practical bandwidth ~25 GB/s per direction
One-way transfer 14.7 MB ÷ 25 GB/s ≈ 0.59 ms
Round trip (GPU-1 → GPU-2 → GPU-1) ~1.18 ms
TLP headers, credit-based flow control, sync ~0.1–0.3 ms
Total added latency ~1.3–1.5 ms
This is on top of the latency the frame already had. The DLSS processing time itself doesn't add new latency — it was always going to happen — but it now sits between two PCIe transfers on the critical path.
It was still missing something so added some thoughts about the overall expected performance cost of the mod itself. That changed things quite considerably.
Quote:
There are several layers of cost beyond the raw PCIe transfer that I under-weighted:
1. Driver Hook / Interception Overhead
The modified driver must intercept the normal render→DLSS→present pipeline. Every frame, the CPU-side driver code now does additional work.
Signal render complete on GPU-1 (fence/event)
Queue the outbound transfer to GPU-2
Schedule the DLSS kernel on GPU-2's command queue
Wait for GPU-2 completion (poll or interrupt)
Queue the return transfer
Signal present to the display
2. Cross-Device Synchronization Cost
The multi-GPU synchronization research (NVIDIA V100/P100 study) shows that cross-device sync over PCIe is significantly more expensive than intra-device sync. Even with hardware-assisted P2P, each synchronization point (fence signal → wait → release) carries:
Ring buffer polling on the CPU or GPU side
Completion signaling (interrupt or MSI-X)
Cache coherency flushes if the frame metadata crosses the boundary
3. The "Modified Driver" Quality Factor — This Is the Big One
This is where your instinct is strongest. A community-modified driver (like the DLSS5-Feeder project on GitHub) introduces costs that NVIDIA's own driver would never have.
Memory allocation Stock driver uses pinned, pre-allocated, cache-line-aligned buffers. A mod may use generic allocations that aren't zero-copy, adding a CPU memcpy hop
Command batching Stock driver batches hundreds of draw calls per submission. A mod that injects per-frame sync points fragments the batch, increasing submission frequency
Polling vs. interrupts A well-written mod uses GPU-side fence polling (no CPU wake). A less-optimized one polls from the CPU, burning a core and adding jitter
Ring buffer management Two GPUs means two ring buffers the CPU must keep filled. If the mod doesn't pre-allocate deep enough rings, the CPU stalls waiting for GPU-2 to drain before it can submit the next frame
Error handling / fallback paths A mod that checks for GPU-2 availability, thermal state, or fallback-to-single-GPU every frame adds branch overhead to the hot path
4. Revised Total Latency Budget
Component Time
PCIe round trip (transfer) ~1.2 ms
Driver hook / extra API calls ~0.3–0.5 ms
Cross-device sync (fences, interrupts) ~0.2–0.4 ms
Pipelined handoff (1-frame delay) ~6.9 ms (at 144 FPS)
Total added latency ~8.6–9.0 ms
This aligns with Digital Foundry's measurement of the DLSS 5 dual-GPU mod adding 8–12 ms to the pipeline. The PCIe transfer is the small part — the synchronization architecture and the pipelining requirement dominate.
It's an added latency of at LEAST ~1+ frame(s) in a game rendering at 1440p @ 144 Hz just from the process of the mod itself as far as deriving a plausible penalty for the whole process.
It's not really much of an issue though. The main benefit is that GPU-1 can render higher in-game quality settings at a good framerate since it doesn't have to deal with the DLSS5 processing.
* - The reason I chose the second cheapest 9070 XT is because the cheapest (€680) is the Acer Predator BiFrost Radeon RX 9070 XT 16GB OC which is plagued with various problems: build quality issues, coil whine, crappy and noisy cooling solution leading to record high hot spot and VRAM temps which in turn leads to serious thermal throttling (~15-20% lower clocks due to power limitation).
These issues can happen with any card, though I suppose some are more prone than others. How often do I read about people getting a card of brand x with issues and then saying they will never buy a card from that brand again. Well it happens to all of them.
Now of course the cooling solution matters. Coil whine can happen to all the cards of any brand. Sometimes you get it sometimes you don't. Sometimes the thermal pad doesn't sit right and that makes for a huge Hotspot temp. And guess what, Nvidia doesn't even give you these temps so you will never know.
And funnily, the Predator Bifrost is actually their higher tier card, allowing up to 340W TDP OoB, meaning it has a thicker cooler and probably the exact same fans as the Nitro which is rated at 304W TDP. No wonder it gets hot, 340W is pushing these cards to their absolute limit. (useless if you ask me) Not that it makes a huge difference in FPS. I'd probably get the Bifrost and reduce the PL to 300W, do some UV and be done with it. With that huge cooler, it will probably run more silent.
Anyway, I removed the fans on my ASUS Prime and put some Noctua fans on it with some UV.
"Enlightenment is man's emergence from his self-imposed nonage. Nonage is the inability to use one's own understanding without another's guidance. This nonage is self-imposed if its cause lies not in lack of understanding but in indecision and lack of courage to use one's own mind without another's guidance. Dare to know! (Sapere aude.) "Have the courage to use your own understanding," is therefore the motto of the enlightenment."
Ok, seems to work but if I'm understanding this correctly some of the details are insane, especially if you have an AMDerp card (big shock, right? )
It *does* seem to work, at least in Oblivion R the common imperial city to chorrol run @ 250 speed (meep meep ) seemed to go smoother BUT OBR alone was gobbling up 15GB of that cache so guess it's great if you're just playing ONE single problematic UE5 game.
Gave Silent Hill f a shot and the game is actually pretty enjoyable so far but if the math is right, some of that OBR cache already went out of the window:
Or maybe I'm getting this backwards and the oblivion data would have been pushed out by the SHf data de to the AMDerp cache limit if this program wasn't saving it on disk?
Interesting stuff mens
Sad that it doesn't seem to work for lords of the fallen 2023 though:
Which is a bit dumb since in order to use the good shit like Optiscaler and UltraPlus you're already bypassing that easyanticheat crapopla anyway.
Guess Elden Ring is probably also out of the question too, but I'd love to see what this could do in the earlier versions of that before they nerfed the draw distance a bit to improve performance...
boundle (thoughts on cracking AITD) wrote:
i guess thouth if without a legit key the installation was rolling back we are all fucking then
I've tried it on a couple of UE games that don't have any recompiling at start-up (ACC and Ready or Not) and it worked fairly well. The program recognizes some Unity games too (i.e. Rogue Trader) as long as they don't use obsolete versions of the engine.
I suppose it won't do much for the extra-pesky traversal stutters since those are engine-coded so Dead Space Remake, SH2 Remake, Oblivion Remustard etc. will likely still have issues in specific spots. Still, it's something
Put.. put a 9070XT bios on it. (if possible with your card)
(not financial advice)
"Enlightenment is man's emergence from his self-imposed nonage. Nonage is the inability to use one's own understanding without another's guidance. This nonage is self-imposed if its cause lies not in lack of understanding but in indecision and lack of courage to use one's own mind without another's guidance. Dare to know! (Sapere aude.) "Have the courage to use your own understanding," is therefore the motto of the enlightenment."
Put.. put a 9070XT bios on it. (if possible with your card)
(not financial advice)
Yeah, should be possible.
Doubt its needed at the time, still faffing about on 1080p (monitor needs an upgrade) and whatever I throw at it except path tracing it does so without much effort.
Ok, seems to work but if I'm understanding this correctly some of the details are insane, especially if you have an AMDerp card (big shock, right? )
Finally gave it a go on the gayming laptop (RTX3060M) and it's yet another AMDerp L confirmed:
Though the real losers here are UE5, lazy/inexperienced derpelopers and us (if you're paying 50/60/70€ for a game you should seek psychiatric evaluation).
Now also need to test transferring the Silent Hill f data from the desktop to the laptop to see if the program feeds it directly to the shader cache (from AMDerp desktop to Jewvidia Laptop) to continue this wacky science experiment.
Morphineus wrote:
Did undervolt it.
What did you use and is there a tutorial?
boundle (thoughts on cracking AITD) wrote:
i guess thouth if without a legit key the installation was rolling back we are all fucking then
Didn't really use an in-depth tutorial just saw it in a HW unboxed video when I was deciding what card to get.
Ended up stable at:
Voltage Offset: -100mV (started at 80 and went up from there)
Max Freq.: 2718MHz (didn't try to push this further)
Power limit: +10%
In adrenaline, if there's a better way let me know. Was the first time for me doing this on a GPU.
Nice, thanks men.
Never really did go that deep on this adrenalin control panel, didn't know there was a tuning tab.
Did two successful 10min tests at -25mV and then -50mV, now need to pick STALKER2 back up in order to give this undervolt some proper trashing to see if it holds or not. Power limit I'll leave alone for now.
And welcome back, even if it's just for a while
boundle (thoughts on cracking AITD) wrote:
i guess thouth if without a legit key the installation was rolling back we are all fucking then
Hehe thanks, yeah I'll jump in now and then. Was lurking again a bit more lately.
I did some benchmarking for a couple hours and then went to test it out in games. The ones I've used: Oblivion the soulless version, Hellraiser (pretty much been playing it the past week).
No real reason to why these, just wanted some longer gaming sessions to see if it stayed stable in those conditions.
And then just: Cyberpunk, Witcher Remake just to compare performance with the RTX 2070 I had before as I ran those two games on it.
Signature/Avatar nuking: none (can be changed in your profile)
You cannot post new topics in this forum You cannot reply to topics in this forum You cannot edit your posts in this forum You cannot delete your posts in this forum You cannot vote in polls in this forum
)
)
)
