3DMark DirectX Raytracing Feature Test Benchmarks

It not ready in general period if it not full RT is not RT and peformance hit is to huge and that on card that cost how much?


Not gona see anything in game anytime soon in true form, We all remember how heavy the "RT" was in Quake II RTX a game that would run 500+ fps on current day cards and probably 4k without RT and probably 1000+ at 1080p with out RT but RT on that 1080p was run barely past 60 fps and that wasnt even true RT

RT is Marketing name at this point still even the console "used" RT as means to sell it

 

No doubt it's being used as a marketing tool - as pretty much any graphics technology would be.

But it seems pretty clear that going forward we are moving towards full ray tracing... whether we get there in 10 years or 20... that's where we're going.

 

We are getting there. Without existing RT hardware we would not have two full RT games (Minecraft, RTX Quake 2).

 

I been hearing bout RT for years BEFORE 2xxx "introduced" I been hearing about OLED more then 2 decades and it still something not widely available due to price. I like the IDEA of RT but anything but ready

 

Its funny how nvidia claimed 2x more ray tracing performance , second generation etc... YET there is ZERO PERCENT improvement .

 

I fundamentally disagree with this and thus the rest of your post isn't really relevant unless you can convince me otherwise.

There's tons of examples of feature sets used in limited cases in order to improve graphics. For example SVOGI used for GI in various games can also be used to do reflections, shadows, etc but mostly isn't. No one sits there and says "well game isn't fully using SVOGI so it's not really SVOGI".

You cast a single ray, you're using raytracing. You cast a few thousand rays and now you can build a reflection model that's more accurate than any previous nonRT implementation. The question becomes do you have the hardware to do that in a performant way? The answer is now yes.

To be clear they didn't claim that. They claimed RT cores were 1.7x faster but RT workloads aren't entirely ran on RT cores, they touch various other parts of the chip which might not necessarily be 1.7x faster.

Further they showed this slide:

https://hexus.net/media/uploaded/2020/9/1c6c4b25-1d6c-480f-beb8-8d41fc3ce581.jpg

Which clearly shows a number of instances where it's not 2x.

 

Result with my Asus TUF 3090..

 

Is not Nvidia Hopper and AMD RDNA3 rumored to be first generation GPUs to use
MCM design kinda similar to current AMD CPUs?
Kinda makes sense if they do, so they can make more smaller chips faster and
cheaper with less failed ones like AMD is currently is doing with Ryzen, Threadripper
and Epyc.

If so it is going to be interesting to see how performance and RT stuff scale hopefully
without the penalties and the need for multiGPU profiles that multiGPU systems
currently have.

Maybe something like
RTX 4060Ti / RX 7700 (2 core)
RTX 4070 / RX 7800 (4 core)
RTX 4080 / RX 7800XT (6 core)
RTX 4090 / RX 7900XT (8 core)
with something similar to AMDs Infinity Fabric connecting the cores.

 

I could be wrong but believe Intel will likely have the first GPU's designed using MCM architecture.

 

I am not sure either but it would make sense for Intel to go for a MCM design,
more so if AMD and Nvidia is doing the same for their next gen.

 

Got what I expected hahaha, 34.14FPS on my RTX 2080Ti Strix,probably will be able do 35FPS but that's it

Denoise is nice but you literally loosing lots of details like normal maps or bump maps etc, I prefer not to use it in pro works

Hope this helps

Thanks, Jura

 

, The processor …..yes I cannot wait , only few days to go .

 

Like others said, this test is the blandest, most unimaginative thing I've ever seen 3DMark release. UL is probably to blame.

 

Those 3090 scores are damn impressive, anything RT related the 30 series shine.

 

Problem is in transistor count. Currently it is beyond what any forge can make at reasonable price for consumer enthusiast.
We'll need maybe 5nm maybe 3nm. But definitely we will need chiplets. And I have feeling that chiplets will have like 99% performance scaling on such workload.

@tsunami231 : Benchmark does not use static image, it uses static scene. That's because people need something to look at. In transition from one camera view to another, you can see that even 20 rays per pixel is far from enough.
And when I imagine how transparent, translucent and dust like objects would "artifact" from not enough rays/bounces...

I think that before we get Full raytracing, we'll get temporal-aliased raytracing. Where each frame GPU calculates amount of rays it can within limited frametime, and then result will get processed with result from last frame. And this result will be baked into special surface textures on each object. (To be used next frame again. Maybe even sequence of more temporal surface textures so there can be weighting based on age of data.)
And that will not only require extra VRAM, but it will require extra memory bandwidth.

@Denial : You know how I look at it. When RT content in game does not provide raytracing like visuals, I do not consider it reasonable improvement. I want visuals to be closer to photorealistic techniques. Not accurate reflections of ugly rasterized scene as that reflection would be nothing more than accurate reflection of ugly objects. No matter how many ray-stickers people place on product, if I see nothing better than traditional rasterized visuals, I'll not be satisfied.

 

btw, i cant understand how they are standardized.....these benchmark.....system....

 

This is quite nice.

 

I think we need someone to explain what is going on in the benchmark for people to actually see it as impressive. Because it really is impressive what's going on. We have actual focal depth changes based on physicality, not someone plopping down a blur filter on the background.

 

CPU: 5900XT 5GHz w/4000MHz DDR4 CL14 (We are ready for Next-Gen)