GART: Games and Applications using RayTracing

[Frenetic Pony]

Yeah it is terrible, but terrible as in.. it would never be terrible if you were to use those actual resolutions.

Imgsli

Imgsli

Its amazing how trickery stuff can give you big perforamnce alongside with "presentable" image. I mean look at this, they both perform similary, yet one looks leagues ahead.

Personally, I've managed to run the default config (with 3 bounces, RTDI and RESTIR on), 1440p dlss performance at around 40-45 FPS on my 3070. I find the whole "it is unplayable, no card can run this" as hyperbole.

1440p dlss perf with 3 bounces;

ofc I would like to run this game at 4k/dlss perf at least, but at this point I can't ask for more. its a 3070. tbh, its not meant for path tracing but I'm grateful it can run it at all.

That is indeed the good news! While Nvidia might not care about making their tech demo stuff runnable, it doesn't mean users can't make it runnable. I wonder what settings tweaks will be available for The Witcher 3 next gen update.

 

[Dampf]

Yeah it is terrible, but terrible as in.. it would never be terrible if you were to use those actual resolutions.

Imgsli

Imgsli

Its amazing how trickery stuff can give you big perforamnce alongside with "presentable" image. I mean look at this, they both perform similary, yet one looks leagues ahead.

Personally, I've managed to run the default config (with 3 bounces, RTDI and RESTIR on), 1440p dlss performance at around 40-45 FPS on my 3070. I find the whole "it is unplayable, no card can run this" as hyperbole.

1440p dlss perf with 3 bounces;

ofc I would like to run this game at 4k/dlss perf at least, but at this point I can't ask for more. its a 3070. tbh, its not meant for path tracing but I'm grateful it can run it at all.

Could you make comparisons with RTXDI on and off in the developer settings? I struggle to tell the difference and it gives a huge performance boost. That's surely a lot less noticeable than using a single bounce or DLSS Ultra-Performance. I am surprised no one mentions this easy performance win anywhere.

 

[techuse]

https://twitter.com/x/status/1601036292380250112

Did anyone expect an Nvidia project to bother with any AMD optimizations? Using vulkan to intercept DX9 draw calls is already probably hugely inefficient and hacky to begin with. It's unreasonable to expect them to bother with performance tuning on competitor GPUs tbh.

 

[Malo]

Did anyone expect an Nvidia project to bother with any AMD optimizations? Using vulkan to intercept DX9 draw calls is already probably hugely inefficient and hacky to begin with. It's unreasonable to expect them to bother with performance tuning on competitor GPUs tbh.

I got the impression that DXVK is rather well optimized.

 

[techuse]

I got the impression that DXVK is rather well optimized.

I would think this is different since this is not just translating code through a wrapper but replacing draw calls.

 

[Kaotik]

I got the impression that DXVK is rather well optimized.

DXVK is, but like techuse said, it's not just translation in case of RTX Remix, so DXVK doesn't really apply here

 

[OlegSH]

Did anyone expect an Nvidia project to bother with any AMD optimizations?

Guess this ubershader doing bad on AMD is exactly in line with what I said before: ""uber-materials" cause low occupancy and divergent shading, that's why inlining imposes restrictions in practice"

It obviously includes tons of TraceRay() calls. Given that these calls are mostly for divergent secondary bounce rays, low occupancy is something I would expect on AMD hw with SW traversal.

Even 4090 spends a lion portion of time in the TraceRay() and it doesn't suffer from the low occupancy issue since traversal is in HW, and there must be an order of magnitude difference or more between something like 4090 with SER and OMMs vs 6900 XT, so 99 ms spent in TraceRay() calls on 6900 XT sounds pretty realistic to me.

 

[trinibwoy]

Does anyone know how many paths Portal RTX updates per frame for either DI or GI? Is it a full resolution pass?

 

[Kaotik]

Even 4090 spends a lion portion of time in the TraceRay() and it doesn't suffer from the low occupancy issue since traversal is in HW, and there must be an order of magnitude difference or more between something like 4090 with SER and OMMs vs 6900 XT, so 99 ms spent in TraceRay() calls on 6900 XT sounds pretty realistic to me.

There can be order of magnitude difference between 4090 and 6900, but even RTX 3060 is beating 6900 silly in this, probably even RTX 3050 if someone bothered to bench it

 

[OlegSH]

probably even RTX 3050 if someone bothered to bench it

Probably because they don't hit occupancy wall. It will be interesting to see whether RDNA 3 will hit the same wall with 1.5x more registers per thread and likely more wavefronts in flight (don't remember whether it was explicitly mentioned in slides, but more rays in flight implies that). Higher RDNA 3's RT performance with more rays in flight goes pretty much in hand with register file size adjustments, which makes a perfect sense given how rays are processed in RDNA, so more register may provide higher than expected scaling for RDNA 3 here (or may not if they will hit the same wall with more registers).

 

[Dampf]

Portal RTX RTXDI on vs off comparisons

RTXDI on:

https://imgur.com/yMZ4tpH

RTXDI off:

https://imgur.com/wN1cJA3

RTXDI on:

https://imgur.com/9wGeos9

RTXDI off:

https://imgur.com/twRQvLe

 

[trinibwoy]

What does turning off RTXDI mean? Is there a fallback path for rasterized primary visibility?

 

[TopSpoiler]

There are not many notable direct light source in this game. Energy orb and the backrooms(?) are the good light source.

Direct Lignting off, RTXDI off, ReSTIR GI off


Direct Lignting on, RTXDI off, ReSTIR GI off


Direct Lignting on, RTXDI on, ReSTIR GI off


Direct Lignting on, RTXDI on, ReSTIR GI on


Direct Lignting on, RTXDI off, ReSTIR GI on

 

[trinibwoy]

Oh wow so RTXDI isn’t actually doing primary visibility? That’s disappointing.

 

[TopSpoiler]

Oh wow so RTXDI isn’t actually doing primary visibility? That’s disappointing.

If I understood correctly, RTXDI is the set of functions for caching and reuse light samples over frames and nearby pixels using ReSTIR. So itself doesn't do any ray tracing tasks like TraceRay.

 

[DegustatoR]

Oh wow so RTXDI isn’t actually doing primary visibility? That’s disappointing.

It's Direct Illumination isn't it? Primary visibility is still rasterized probably? Q2RTX was the only game where you could use RT for primary rays IIRC.

 

[Dampf]

There are not many notable direct light source in this game. Energy orb and the backrooms(?) are the good light source.

Direct Lignting off, RTXDI off, ReSTIR GI off
View attachment 7752

Direct Lignting on, RTXDI off, ReSTIR GI off
View attachment 7753

Direct Lignting on, RTXDI on, ReSTIR GI off
View attachment 7754

Direct Lignting on, RTXDI on, ReSTIR GI on
View attachment 7755

Direct Lignting on, RTXDI off, ReSTIR GI on
View attachment 7756

That's a bigger difference. But considering the performance boost you get, it's a good tradeoff. Light balls still cast dynamic shadows and lighting with RTXDI off.

 

[Dictator]

Primary visibility is done via RT - Turing Off direct lighting means you are turning Off the first hit contribution but still showing all other bounces lol

Not possible in real life, but sure possible in ray tracing!

It is useful for debugging and seeing the contribution of light bounce in an image.

You can actually do that in quake2RTX with the debug presets available.

 

[OlegSH]

So itself doesn't do any ray tracing tasks like TraceRay.

How would it produce shadows without tracing rays? RTXDI is responsible for direct lighting, so it should cast quite a few shadow rays)

 

[trinibwoy]

Primary visibility is done via RT - Turing Off direct lighting means you are turning Off the first hit contribution but still showing all other bounces lol

Not possible in real life, but sure possible in ray tracing!

It is useful for debugging and seeing the contribution of light bounce in an image.

You can actually do that in quake2RTX with the debug presets available.

The first pic above from @TopSpoiler has Direct Lighting, RTXDI and RestirGI all turned off. I interpret that to mean some basic rasterized lighting is still being performed.

RTXDI needs information about the material properties for each pixel which is basically a rasterized g-buffer. Presumably you can also generate that buffer using RT primary rays. It looks like Portal RTX doesn’t actually need the light contribution from raytraced direct or indirect illumination to illuminate its pixels.