Digital Foundry || Ray Tracing Performance Deep Dive: AMD Radeon RX 6800 XT vs Nvidia GeForce RTX 3080

[ILikeFeet]

get ready for the world's most one-sided fist-fight. but I already drank the RT kool-aid many years ago, so I'm interested regardless

Many thanks to XFX for the RX 6800 XT Speedster Merc 319 Black sample. Yes, we've seen the benchmarks - and it's undeniable that Nvidia's second generation ray tracing technology is faster than AMD's debut offering in the BIG NAVI line-up of RDNA 2 graphics cards. But we wanted to know more - what are the strengths and weaknesses of the respective architectures? Is Nvidia really a full generation ahead of AMD performance or is the answer a little more complex? And how does this all break down when considering different forms of ray tracing? Alex Battaglia loves ray tracing, loves telling you about it - and he has everything you need to know.

Click to shrink...

written article

PC ray tracing deep dive: Radeon RX 6800 XT vs GeForce RTX 3080

AMD's brand new RDNA 2 architecture has arrived for desktop PCs via the RX 6000 line of graphics cards - and it's an im…

www.eurogamer.net

 

[dgrdsv]

I thought that the results were interesting in the part where increasing the RT workload in number of rays shot would lead to smaller wins of 3080 over 6800XT than by just turning RT on.

Thinking on that for a bit I tend to presume that the more rays are being shot the more of a bottleneck the BVH traversal step become, to a point where Ampere's advantages in pure compute (flops) and hit evaluations (dedicated MIMDs) becomes less relevant as the actual performance limiting part becomes the ray traversal which is done by RT h/w on both Ampere and RDNA2.

 

[Ambient]

I want a 6800 but I'm not paying scalper prices.

 

[Stacey]

I would have liked a video that explains the difference in laymans terms, frame rate.

Also, Dirt 5 has RT shadows? I just got done finishing the single player and didnt notice an option, it surely isnt baked in?

 

[ILikeFeet]

I would have liked a video that explains the difference in laymans terms, frame rate.

Also, Dirt 5 has RT shadows? I just got done finishing the single player and didnt notice an option, it surely isnt baked in?

Click to shrink...

I don't know if the game has been updated, but RT shadows was on a special build that only reviewers had

 

[brawndolicious]

It sounds like most of the cost of RT in these games comes from the initial collision testing and the final denoising. With the roughness cutoff directly increasing the cost of the initial and final steps (IE: more ray traceable surface area to deal with). Assuming my brain didn't mix up the numbers.

I get that this is the way to do it for conventional RT, but it sounds like there's going to be a lot of motivation to come up with new methods that get around that heavy initial/final step performance cost. Maybe by avoiding recalculating every single reflective surface by just checking if the lighting around it has stayed the same?

 

[RoboPlato]

It sounds like most of the cost of RT in these games comes from the initial collision testing and the final denoising. With the roughness cutoff directly increasing the cost of the initial and final steps (IE: more ray traceable surface area to deal with). Assuming my brain didn't mix up the numbers.

I get that this is the way to do it for conventional RT, but it sounds like there's going to be a lot of motivation to come up with new methods that get around that heavy initial/final step performance cost. Maybe by avoiding recalculating every single reflective surface by just checking if the lighting around it has stayed the same?

Click to shrink...

It'll be fascinating to see how RT evolves over the next few years now that consoles support it and more and more PC GPUs will have native support. Lots of space to grow in different ways in small period of time.

 

[ILikeFeet]

It sounds like most of the cost of RT in these games comes from the initial collision testing and the final denoising. With the roughness cutoff directly increasing the cost of the initial and final steps (IE: more ray traceable surface area to deal with). Assuming my brain didn't mix up the numbers.

I get that this is the way to do it for conventional RT, but it sounds like there's going to be a lot of motivation to come up with new methods that get around that heavy initial/final step performance cost. Maybe by avoiding recalculating every single reflective surface by just checking if the lighting around it has stayed the same?

Click to shrink...

I remember hearing that Nvidia has been working to get denoising off the shader cores and onto the tensor cores, something I think is already available to offline renderers. would be a boon for performance, especially for stuff like Minecraft

 

[KDR_11k]

I remember hearing that Nvidia has been working to get denoising off the shader cores and onto the tensor cores, something I think is already available to offline renderers. would be a boon for performance, especially for stuff like Minecraft

Click to shrink...

Wouldn't that get in the way of DLSS?

 

[ILikeFeet]

Wouldn't that get in the way of DLSS?

Click to shrink...

DLSS happens after that, so it shouldn't

 

[KDR_11k]

DLSS happens after that, so it shouldn't

Click to shrink...

Hm, I was almost thinking that it'd be pipelined, i.e. while the DLSS does its thing the next frame already starts rendering in the other parts of the card.

 

[Jamrock User]

I wish he would also add a fps counter.

I get his reasoning but it made the video so much harder to follow.

 

[dgrdsv]

Wouldn't that get in the way of DLSS?

Click to shrink...

It would get in the way of one RT code path working on all vendors including those who don't have tensor cores.
Although if implement via DirectML it may possibly work.

 

[Vulcano's Assistant]

I wonder by how much the relative cost of that second GI bounce would tank in the 2080ti.

 

[Pipyakas]

I wish he would also add a fps counter.

I get his reasoning but it made the video so much harder to follow.

Click to shrink...

If the frame time is lower than 16ms, it's better than 60fps
If the frame time is higher than 33ms, it's worse than 30fps
I honestly dont know what more should be said about frame rate, as Alex said the actual performance difference is not the point and namely has been done by every other RT-related reviews or content

 

[Pottuvoi]

It sounds like most of the cost of RT in these games comes from the initial collision testing and the final denoising. With the roughness cutoff directly increasing the cost of the initial and final steps (IE: more ray traceable surface area to deal with). Assuming my brain didn't mix up the numbers.

I get that this is the way to do it for conventional RT, but it sounds like there's going to be a lot of motivation to come up with new methods that get around that heavy initial/final step performance cost. Maybe by avoiding recalculating every single reflective surface by just checking if the lighting around it has stayed the same?

Click to shrink...

Depending on case the shading can cost a lot as well. (And reason why usually shaders used for RT are quite bit simpler, no normal maps etc.)
On UE4 one can test it with r.RayTracing.EnableMaterials, this changes all RT materials to test material.

 

[brain_stew]

I'd really like to congratulate Alex on another fantastic piece of work. This may not get the views of some other content but the level of detail and investigation here is unparalleled.

 

[Serious Sam]

Oh look, two graphics cards that no one can buy.

Nice vid btw.

 

[sweetmini]

This was an interesting video, and an angle not many tried to grapple: the costs of RTX in ms. It's always good to take a different point of view to better get a grasp on the reality.

Pretty good job, and i also hope to see future amd optimized titles, that will use pipelines that would optimize the cache advantage and things like that.
We need proper benchmarks with toggable stages ... sure it's mainly a scholar's interest, but still an interest :P

Thanks for sharing OP.

 

[Hope]

TLDR? Iam at work so can't watch.

 

[ILikeFeet]

TLDR? Iam at work so can't watch.

Click to shrink...

Turning on RT from off to medium incurs a higher penalty on AMD, but as quality increases, Nvidia takes more time to render the RT effects (Nvidia is still way faster however).

AMD also seems to do better with more rays/samples, while Nvidia does better with more complex geometries.

In Watchdogs, Nvidia's Gameworks denoiser works better than the denoiser on consoles, on both AMD and Nvidia

As some olpeople on reddit put it, once Nvidia's RT acceleration hits its limit, it becomes a more drastic bottleneck. AMD doesn't since more is done on the shader cores (but this method also means RT is inherently slower)

 

[Hope]

Turning on RT from off to medium incurs a higher penalty on AMD, but as quality increases, Nvidia takes more time to render the RT effects (Nvidia is still way faster however).

AMD also seems to do better with more rays/samples, while Nvidia does better with more complex geometries.

In Watchdogs, Nvidia's Gameworks denoiser works better than the denoiser on consoles, on both AMD and Nvidia

As some olpeople on reddit put it, once Nvidia's RT acceleration hits its limit, it becomes a more drastic bottleneck. AMD doesn't since more is done on the shader cores (but this method also means RT is inherently slower)

Click to shrink...

Thank you:)

 

[daninthemix]

I found it interesting from one of their other videos, when they explain that turning any one ray tracing feature on incurs the most cost, with additional features / improvements to quality usually costing much less. So any ray tracing at all has the biggest cost.

 

[BreakAtmo]

DLSS happens after that, so it shouldn't

Click to shrink...

This makes me think that when gaming the tensor cores must kind of sit around doing nothing until they're called up to do DLSS. I look forward to the point where enough AI-based features are implemented that they're working all the time.

 

[ILikeFeet]

This makes me think that when gaming the tensor cores must kind of sit around doing nothing until they're called up to do DLSS. I look forward to the point where enough AI-based features are implemented that they're working all the time.

Click to shrink...

Well, dlss can also work concurrently as the shader cores are rendering another frame, but it's still a relatively light load on the tensor cores.

 

[Deleted member 27751]

I've only tested RT in Deliver Us the Moon with my 6800 XT so I do have limited benchmarks, but it is surprising just how useless RT's graphical impact is with the performance hit. I get roughly 120-144 FPS without RT on all ultra, when turning RT on it does drop to around 80 FPS (freesync makes it non-existent) and all I get is some prettier light castings. Granted the game is already phenomenal in the art department so that probably helps to negate any "boon" RT would give, but I'm really happy that I even have the option on an AMD card. Either way it is a massive improvement on my 1070 to be able to reach 144 FPS @ 1440p.

 

[BreakAtmo]

Well, dlss can also work concurrently as the shader cores are rendering another frame, but it's still a relatively light load on the tensor cores.

Click to shrink...

Do you know what could be done to push them harder? Right now we have DLSS, soon we should have AI denoising. Are there any other applications in the works that you know of?

 

[icecold1983]

Turning on RT from off to medium incurs a higher penalty on AMD, but as quality increases, Nvidia takes more time to render the RT effects (Nvidia is still way faster however).

AMD also seems to do better with more rays/samples, while Nvidia does better with more complex geometries.

In Watchdogs, Nvidia's Gameworks denoiser works better than the denoiser on consoles, on both AMD and Nvidia

As some olpeople on reddit put it, once Nvidia's RT acceleration hits its limit, it becomes a more drastic bottleneck. AMD doesn't since more is done on the shader cores (but this method also means RT is inherently slower)

Click to shrink...

This isnt accurate. Nvidia doesnt take more time to render higher quality RT effects or do worse with more samples. In the tested scenarios Nvidia is faster at all levels of RT, just less so as the RT load increases. RT incurs a higher performance penalty on AMD at all quality levels.

 

[orava]

Interesting topic and results but man that was a chore to watch.