Yah, I'd agree with both of these. Cooperative shaders needs to launch in dx12 and then it's going to take years to start seeing anything, most likely. There may be a few Nvidia supported titles that get things early, but we'll see. I was surprised they didn't have any announcements about that tech showing up in a particular game, like Doom or something. Tells me it's a ways off.
Yah, I'd agree with both of these. Cooperative shaders needs to launch in dx12 and then it's going to take years to start seeing anything, most likely. There may be a few Nvidia supported titles that get things early, but we'll see. I was surprised they didn't have any announcements about that tech showing up in a particular game, like Doom or something. Tells me it's a ways off.
Hopefully we will see Neural Texture Compression soon. That's the most exciting feature imo. Reducing VRAM usage is a huge, huge deal especially considering this gen is so VRAM starved (with the exception of the 5090.)
But I'm also excited for RTX Mega Geometry, accelerating RT performance is also a big deal. We should expect seeing all RTX cards get uplifts in Alan Wake 2, but Blackwell will likely benefit from it the most. Couple that with the transformer model running better and we should see maybe a 60% uplift in performance, before frame gen, which would be a good gen on gen improvement again.
But I'm also excited for RTX Mega Geometry, accelerating RT performance is also a big deal. We should expect seeing all RTX cards get uplifts in Alan Wake 2, but Blackwell will likely benefit from it the most. Couple that with the transformer model running better and we should see maybe a 60% uplift in performance, before frame gen, which would be a good gen on gen improvement again.
Ok, as promised, I ran some benchies and did a bit of "walking around" in Cyberpunk 2077, specifically to compare the new TNN model against the older CNN model for DLSS upscaling modes. Here's the specs I'm running:
And here are the results:
A few observations:
First, I ran the benches several times, and the numbers there are repeatable on my rig. For whatever reason, TNN Balanced very slightly outperforms CNN Balanced on my setup somehow... It's only a single percentage point or two, but it's a repeatable difference across three different runs of the benchmarks. But then at TNN Performance, it gets wallaped by CNN performance to the tune of almost 30%, however the visible artifacting of CNN in Performance mode makes it a bit of a "meh" comparison at best. IMO, the TNN ultra performance mode is a higher fidelity visual output than the CNN performance mode. If you're on a lower end card, I'd suggest at least trying TNN Ultra Performance...
Second, I agree with the earlier statement that TNN can and should be compared (visually) with the +1 step of CNN -- meaning I found TNN in Balanced mode to be every bit as good as CNN in Quality mode. In fact, outside of two very specific cases, I think TNN Performance was essentially as good as CNN Quality to my eyes. The variations were thin wires and fences, which did a better job of simply fading out versus being an aliased dotted line mess. Given the performance, I'd just stick with TNN Balanced every day.
Finally, TNN balanced is, to my eyes, almost indistinguishable from the native raster. After repeated runs and some walking around the city, I found two issues: one was a thin wire fence in the distance, dark wire against a moderately dark background, where the model was having a hard time picking out the fence itself. Second, I found some stationary shadow crawling / movement on along the sidewalk edge when cars would pass by and "Disrupt" the shadow. All day every day I'd take TNN balanced as my daily driver, as the quality setting didn't seem to "fix" either of those issues for me.
I dunno, I think the TNN is pretty damned good for a first pass. There's still room for improvement, but I'm good with it.
- CPU: 9800X3D at mostly stock settings - FCLK is at 2100 but otherwise no PBO tweaks, no voltage tuning, etc.
- RAM: 96GB of DDR5 at 6400MT/s CL30 with UCLK at 3200MHz.
- Disk: My Steam library is on a 4TB 990Pro, with a 2TB 980Pro as the Windows 11 24H2 boot disk
- GPU: Asus TUF Gaming 4090, clock-locked at 2715MHz@964mV core and 11501 mem.
- Cyberpunk Video: No vsync, no max fps, no HDR, resolution is 3440x1440 on full screen, reflex on + boost
- Cyberpunk Graphics: All sliders at maximum value, all toggles on except: no chromatic aberration, motion blur, film grain, lens flare, or depth of field. Ray Reconstruction enabled, path tracing enabled, no framgen.
And here are the results:
- Native / No DLSS: 37.5 avg, 33.5 min
- CNN Quality: 72.8 avg, 66.6 min
- TNN Quality: 71,.2 avg, 65.2 min
- CNN Balanced: 84.3 avg, 76.7 min
- TNN Balanced: 85.3 avg, 77.8 min
- CNN Performance: 130.6 avg, 90.2 min
- TNN Performance: 101.4 avg, 90.3 min
- TNN Ultra Performance: 140.4 avg, 110.9 min
A few observations:
First, I ran the benches several times, and the numbers there are repeatable on my rig. For whatever reason, TNN Balanced very slightly outperforms CNN Balanced on my setup somehow... It's only a single percentage point or two, but it's a repeatable difference across three different runs of the benchmarks. But then at TNN Performance, it gets wallaped by CNN performance to the tune of almost 30%, however the visible artifacting of CNN in Performance mode makes it a bit of a "meh" comparison at best. IMO, the TNN ultra performance mode is a higher fidelity visual output than the CNN performance mode. If you're on a lower end card, I'd suggest at least trying TNN Ultra Performance...
Second, I agree with the earlier statement that TNN can and should be compared (visually) with the +1 step of CNN -- meaning I found TNN in Balanced mode to be every bit as good as CNN in Quality mode. In fact, outside of two very specific cases, I think TNN Performance was essentially as good as CNN Quality to my eyes. The variations were thin wires and fences, which did a better job of simply fading out versus being an aliased dotted line mess. Given the performance, I'd just stick with TNN Balanced every day.
Finally, TNN balanced is, to my eyes, almost indistinguishable from the native raster. After repeated runs and some walking around the city, I found two issues: one was a thin wire fence in the distance, dark wire against a moderately dark background, where the model was having a hard time picking out the fence itself. Second, I found some stationary shadow crawling / movement on along the sidewalk edge when cars would pass by and "Disrupt" the shadow. All day every day I'd take TNN balanced as my daily driver, as the quality setting didn't seem to "fix" either of those issues for me.
I dunno, I think the TNN is pretty damned good for a first pass. There's still room for improvement, but I'm good with it.
@Albuquerque Interesting. Seems like TNN bottlenecks at DLSS Performance. Guess the tensor cores become the limiter at the settings you were running.
That should be the same with CNN. There are no motion vectors for reflected objects, so it is kind of a worst case. Would be good to test that specific case with CNN vs TNN. I wonder if there's something different about it being an RT reflection vs the animated textures that see TNN improvements.
I always test reflections first...I hope to one day see RT refractions in glass but alas, not yet.
But the image is "better" with TNN than CNN, that much I can say from just 15 minuttes of testing.
@IQandHDR I'll check that out later today. I did go looking for blurring or ghosting in movement, but I guess I didn't check reflections for that.
@Scott_Arm your guess is as good as mine; I'm not sure why the difference was so stark at the Performance preset. What I found interesting was how performant the other TNN modes were as compared to the CNN modes, at least given how we saw them described in the links above. With more time, I'll try downclocking my VRAM to see if maybe it's linked to memory bandwidth? And maybe that's at least one reason why the 5090 has such an insanely large increase in memory bandwidth?
@Scott_Arm your guess is as good as mine; I'm not sure why the difference was so stark at the Performance preset. What I found interesting was how performant the other TNN modes were as compared to the CNN modes, at least given how we saw them described in the links above. With more time, I'll try downclocking my VRAM to see if maybe it's linked to memory bandwidth? And maybe that's at least one reason why the 5090 has such an insanely large increase in memory bandwidth?
I find the area around Corpo Plaza to be a good place to test.
Lots of glass, reflections etc.
There I spotted that a reflection in one plane of glass...dissapears if "refracted" by another plane of glass:
Or that the big "flying fish of lights" are only screen-spaced, not RT/PT
Lots of stuff to play with in that area
Flappy Pannus
Veteran
The performance uplift is definitely "med", DLSS Transformer is definitely a great win.
NVIDIA GeForce RTX 5090 Founders Edition Review - The New Flagship
NVIDIA's GeForce RTX 5090 is the fastest graphics card ever released. It comes with a whopping 32 GB VRAM, and support for multi-frame generation, which achieves hundreds of FPS easily. We also managed to disassemble the card, and our review includes performance testing without upscaling or...www.techpowerup.com
Winning over a DLSS cynic is something, wow.
techpowerup said:
You said the uplift was 35-50% making the average sound somewhere in the low to mid 40's. Except the average was literally the minimum you mentioned.
That's a pretty noteworthy difference if somebody is just reading your post and not checking the source to see if that's correct. Actual range is lower than you claimed.
DegustatoR
Legend
It's funny because in most scenes of most games the difference isn't that big, and in some cases there are what you could call "regressions" although these are generally more of a result of bad DLSS implementation in the first place.
PC gamers spent so long glorifying the unerring superiority of 'native resolution', at least since the LED age. Gamers used to sooner drop settings to low or just take the performance hit before thinking about touching resolution.
We always get some folks like this that are just slow to adapt to a new paradigm once an old one they considered basically set in stone has been shattered.
We always get some folks like this that are just slow to adapt to a new paradigm once an old one they considered basically set in stone has been shattered.
I did not. Please read again.