Optimizations of PC Titles *spawn*

Every GCN optimized game runs superb on Ampere. I think it makes sense because those games are compute oriented and focused less on the fixed function side (no Tessellation, low geometry etc.).

 

Image quality in this case is literally more art than science. An objective quantitative measure of IQ isn't very useful at all.

 

https://research.nvidia.com/publication/2020-07_FLIP

I watched the related presentation at HPG 2020 recently. It's pretty interesting and I think we may well have escaped the crappy local minima of SSIM and the like.

You can run a "reference" ray trace pass for game engine graphics (which will take some time) to see what the image quality should be, and then use that reference image to assess the quality of the real time ray tracing code.

 

I was under the impression that NVidia gave developers a "Gameworks" DLL to bundle with the game. Maybe that's no longer the case?

Dunno how the NVAPI DLL relates to "Gameworks":

https://developer.nvidia.com/nvapi
https://docs.nvidia.com/gameworks/content/gameworkslibrary/coresdk/nvapi/group__dx.html

Oh...

 

Yeah, I met Tomas at HPC 2018, awesome guy. I wish I've had such a professor at U way back. I'm also friends with the other Thomas from SSIM. Basically my neighbourhood.
There are crazy ideas floating around there and in general considering machine learning, they look like sick magic, but they are real hardcore math and logic with proven boundaries and predictable behaviour and all. It's sad that PR is giving into the lower instinct (or whatever), instead of trying to educate and explain and lift everyone a step up, mentally.

 

Yeah I’m not saying it’s not possible, just that the output won’t be very useful. Such an objective measure for example will give horrendous scores to screen space reflections and cube maps which we already know are quite acceptable. If it’s used as a measure of how many rays are good enough for a given effect it might be somewhat handy. If it’s used as a measure of which of several competing
effects (shadows, ao, reflections) should get a higher ray allocation it’s really useless.

 

How anyone can unironically complain about a perceived lack of optimisation — or intentional performance gimping — for Nvidia cards on AMD sponsored titles, after over a decade of “The Way It’s Meant to be Played” or GameWorks, is honestly baffling.

If you want to talk about needless compute for no IQ gains, take a look at the tesselation shenanigans Nvidia was pulling just to penalise AMD’s GCN uarch.

 

yeah and in the meantime, one man studio game following MS DX12 RT API, without help of anybody, produces that:


Edit: sarcasm aside, looking at Minecraft RTX or DF analysis of Watch Dog Legion Console vs PC (where XBsX quality is equivalent of what a RTX2060 can do), it's very clear that RDNA2 is far behind RTX30 when talking RT. Nothing to be ashamed for AMD, it's their first generation, but please stop this non sense of gimped titles, that's not what's happening this gen...

 

Are you really comparing Nvidia performance with DLSS to AMD performance at native?

 

Nvidia PR is already doing wonders.
OMG nvidia is 3x faster!!

 

I checked out that video and it was interesting, thanks (Link if anyone's interested). To note, it was a RTX2060 Super they were using, not a vanilla 2060. And it was compared to a 52CU RDNA2 GPU at 1.8 Ghz with no infinity cache. A 72/80CU 2.2 Ghz RDNA2 GPU with Infinity cache will likely perform better. From preliminary tests/blurbs/news,etc it appears at least at RTX 3070 levels, if not higher (Game optimization/driver maturity dependent of course). It seems fairly clear that Nvidia is ahead in Ray Tracing performance, and given it's the second gen of their marquee feature, I don't think anyone's surprised. However, this is still early and there are significant optimizations possible on both sides for various games. So give it some time and wait for actual results instead of jumping to conclusions. If RT is a priority for you and you want to buy RTX cards go ahead, but this incessant noise in literally every thread that AMD's raytracing is bad/horrible/whatever is getting quite ridiculous frankly.

 

Hey, it seems RDNA2 finally shows its true power in Ray Tracing. And look at the SAM gains!
Thanks for optimizations, AMD. Hell, AMD has optimized it sooo well that even RTX 2080 Ti is capable of beating that 3090!

It seems Valhalla and The Dirt 5 optimizations are really portable and help to leverage the potential of RDNA to full extend!

 

Also, why is there this general assumption that Navi 21's RT performance will be stagnant and equal to the performance on release day?

When Turing released, RT performance on the first implementations was pretty terrible while IQ gains were almost negligible.
But AMD releases a new card that is running RT on games optimized exclusively for nvidia's hardware and we should assume this is the best we'll ever see from said new card?

Why does one get the benefit of the doubt for years, while the other gets a lifetime sentence?
Because a couple of developers said nvidia cards are faster at incoherent rays? Perhaps games and game engines should be made to avoid incoherent rays (secondary/tertiary bounces if I understood correctly) as much as possible, since from what I'm reading those are pretty slow on nvidia hardware (and slower on RDNA2, reportedly).
It seems this performance problem has been in the radar from more than a couple of years now.


Unless of course, we're about to see subpixel triangles / Hairworks saga 2.0, which is nvidia pushing for developers to kill the performance on their own cards just because it kills performance on their competition disproportionately more.
And if that's the case, expect to see lots of opposite sided mirrors everywhere just so people can see triple-mirrored rays / images that put the 3080 running at 40 FPS but it's ok because the 6800XT won't go above 10 FPS.

 

The 3DMark feature test should be a best case scenario for AMD. Ampere is twice as fast.

 

Huh? What makes you think it's best case scenario for AMD?

 

The second sentence in the post you quoted has the answer to your question.

 

How exactly "Ampere is twice as fast" best case scenario for AMD? I'm pretty sure 3080 isn't beating 6800XT in most RT games 2:1

 

It's not a statement, it's a will.
(it was a tongue-in-cheek comment on my side)

Just to be clear: I think Ampere (GA102) being "at least twice as fast" at RT than Navi 21 is just what nvidia's PR will try to sell everywhere from now on, since they just got significantly behind on rasterization perf/watt. Their message will change radically from "faster at gaming" to "faster at raytracing" and all of a sudden raytracing will be the most important thing to ever happen in gaming.

 

You’re assuming that faster at gaming and faster at raytracing are going to be 2 different things this generation from a typical gamer’s perspective. Nvidia’s marketing team has a pretty easy job if big titles continue to embrace RT.

Practically speaking Nvidia has no reason to change course. AMD has a very compelling product but it’s not the first time that’s happened and we know how that usually goes in terms of market penetration. All the other stuff that helps Nvidia stay ahead - brand recognition, software, feature set, developer relations are still there. Of course if Nvidia’s advantage in RT solidifies they will beat us over the head with it but that’s expected.

 

I think this is exactly how this and the next generation will play out.

64x tesselation for under water assets, anyone?