I think a long last gen followed by a prolonged period of cross gen has given people unrealistic expectations about how current hardware will run real next gen targetted games.
Current Generation Games Analysis Technical Discussion [2023] [XBSX|S, PS5, PC]
- Thread starter BRiT
- Start date
- Status
I think a long last gen followed by a prolonged period of cross gen has given people unrealistic expectations about how current hardware will run real next gen targetted games.
A 4090 has ~85 TFLOPs compute performance or 4x more than a 6900XT. I think this tells us that software is just unoptimized.
As long as you want to do a desert, it may be right. But when you want to do more an unified lighting systems provides much better graphics.
Subtlesnake
Regular
Compute still requires bandwidth to operate. So is the issue the ALU or the lack of bandwidth to feed the ALU? Does it have 4X more bandwidth?
Secondly, you're redoing lighting and shadow calculations every single frame, the higher the resolution the more compute and bandwidth required.
You're seeing loads never done in previous generation games.
Last edited:
Subtlesnake
Regular
The point is that the level of geometric detail required to render at the level of individual leaves, presents problems for traditional ray tracing approaches. Cyberpunk isn't trying to do that, and looks great.
Have you played through gears 4? I wouldn't blame you if not, it's not the best game ever made, but I think you may have a mis-conception about how it looks from screenshots or clips. It's not as consistent artistically as uncharted or doom, but in it's best scenes it definitely looks just as good.
And the reflection of that is a two triangle block. UE5 cant even do what DICE did with BF5 - five years later.
It just takes additional 10ms to make the impossible possible: https://imgsli.com/MjAyNzIz
What takes the UE5 so long to render a frame without proper direct lighting, reflections, shadows and a worse GI solution on my 4090?!
A 4090 only has 85 theoretical FP32 flops. Due to the nature of how Nvidia does its 2xFP32, it is obvious that there will never be a situation where you get anywhere close to 2x the performance from such a solution since games still need decent use of INT32 instructions. It was not some magic way to actually get twice the compute performance in a game. I think I even looked at this a couple weeks ago here and found that 2xFP32 seems to achieve something like 20-25% better performance in some of the better-case scenarios. Which is still good, but still, nothing remotely close to 100%. This alone cuts the actual difference between these two GPU's massively.
Beyond that, gaming performance generally doesn't scale with TFLOPS like this anyways. You dont really get 30% more performance from 30% more TFLOPS, even with the same general architecture.
It's also a weird argument, cuz you're kind of making the case that this 'unoptimized' factor only applies to the highest end parts? If I can demonstrate relatively predictable scaling from a 6700XT to a 6900XT, what does that then suggest about this supposed lack of optimization?
Subtlesnake
Regular
I was talking about hardware Lumen (I assume the 2nd shot?), since we were comparing with pathtracing. (Reflections with software Lumen have a distance cut off). Alex found with the Matrix demo it had a 7% hit when the CPU was taken out of the equation, but otherwise was heavily bound by single threaded CPU performance. That's also something the Immortals of Aveum developers warned about.
What CPU are you using?
Edit: Ah, it seems a fully path traced version was released on Friday, so you are comparing software Lumen to pathtracing?
Last edited:
Idk, I personally find the look of the non path traced version to be more appealing, aside from the reflection. The blocks appear to be much more grounded and less video-gamey as the shadows around them are better.
This demonstrates nicely what I mean. With path tracing (left) the block appears to be floating over the water while it is grounded with Lumen.
Proves that the more demanding and technically more accurate technique doesn't always result into the better image.
DavidGraham
Veteran
I'd have to totally disagree with that. The PT shot clearly looks way better to me, even my 6 yo commented on "liking the darker one better" without me even needing to ask - and he's usually completely blind to graphics.
In terms of the comparison above, I think it's down to the difference in the water surface which seems to look much better in the PT screen for the most part but is lacking the same intensity of AO beneath those particular blocks there. I expect seeing the image in motion would have an impact there as well.
Full ray tracing also has a much longer range. With enebaled Lumen you can see how the GI and reflections build up in front of you when walking. On the other side full ray tracing is stable.
It's still a bit noisy. Especially when it comes to reflections. But that's where Ray Reconstruction will help.
Lumen has a softer look which may look more natural for some.
It's still a bit noisy. Especially when it comes to reflections. But that's where Ray Reconstruction will help.
Lumen has a softer look which may look more natural for some.
200W for a 4090 is weird, for sure. I know in a lot of games it won't hit its max, but 200W is definitely low.
No, i use hardware Lumen in this game. Looks better.
Here is a comparision between Screenspace GI+Reflections and VSM vs. hardware Lumen vs. Pathtracing with "mat" surface: https://imgsli.com/MjAyNzQy
A few points:
- Without any kind of advanced rendering (except Nanite) the base performance is very slow. With less than 320W the 4090 is on the lower side of power consumption in games. For example Control runs with >100fps with raytracing and 500W. So just 120fps makes it very clear that UE5 is not optimized for nVidia.
- With hardware Lumen performance will drop. And thx to the low base performance HW Lumen puts it down to 86 FPS. The GPU is doing more work (360W) but this is still lower than Control.
- Pathtracing is slower but the power consumption is skyrocketing to 460W. More work, more power consumption. And the content is still the same. And it is the same UE5. Hardware accelerated Pathtracing is actually working in UE5.
Here is another comparision between HW Lumen GI and Pathtracing: https://imgsli.com/MjAyNzQ4
No, the path traced shot looks clearly better lol. It's not even a contest.
Some comparisons from me:
Hardware Lumen vs. Full Ray Tracing 4 at DLSS Ultra Performance (720P): https://imgsli.com/MjAyNzUw
Hardware Lumen vs. Full Ray Tracing 5 at DLSS Ultra Performance (720P): https://imgsli.com/MjAyNzUx
Hardware Lumen vs. Full Ray Tracing 6 at DLSS Ultra Performance (720P): https://imgsli.com/MjAyNzU5
What you don't see in the pictures is that the lumen is also more pulsating and restless while standing still.
Even with DLSS Ultra Performance the image quality is mostly good on a 65 inch screen at 2,7 meters distance. This makes it playable for many more GPUs.
More comparison shots with the DLSS Quality mode (1440p) are here:
Hardware Lumen vs. Full Ray Tracing 4 at DLSS Ultra Performance (720P): https://imgsli.com/MjAyNzUw
Hardware Lumen vs. Full Ray Tracing 5 at DLSS Ultra Performance (720P): https://imgsli.com/MjAyNzUx
Hardware Lumen vs. Full Ray Tracing 6 at DLSS Ultra Performance (720P): https://imgsli.com/MjAyNzU5
What you don't see in the pictures is that the lumen is also more pulsating and restless while standing still.
Even with DLSS Ultra Performance the image quality is mostly good on a 65 inch screen at 2,7 meters distance. This makes it playable for many more GPUs.
More comparison shots with the DLSS Quality mode (1440p) are here:
Last edited:
Man from Atlantis
Veteran
It's AC Valhalla all over again. I'd expect an nvidia driver update that boosts perf around +20% in 6 months.
I can hardly agree with this. If it's the software lumen, then SDF doesn't care about your actual geometry density. if it's hardware RT lumen, then it uses the fallback mesh (not the nanite one, the lowest quality) for raytracing. They actually talked about the self-occlusion problem due to low quality RT mesh in their Matrix tech talk. I don't think geometric density is the issue here.
Last edited:
- Status
Similar threads
