^^
Somebody benched the Star Wars RT demo and was getting 4 fps at 1440p on the 1080 Ti whereas the 2080 was getting like 30ish. I'm sure nVidia didn't do much optimizing for the fallback but I think it's reasonable at this point to assume that you won't really bother using RT without dedicated hardware.
They didn't and it's likely that they purposely gimped it.
Pre-launch, I had discussions on other forums to try to get at the bottom of the claimed Gigaray/sec metrics.
Throughout such discussions, aspects of hardware/software were probed in detail.
In particular, a user grabbed DX12 and probed DXR fallback code.
In it, he found a handful of one line code changes that resulted in performance doubling and it had to do w/ the type of memory primitive that was used. BVH has been stated to be constructed in CUDA cores. RT cores only accelerate intersection tests.
Performance drops through the floor when you have reflections and ray divergence which is why they are toned down even on Geforce20 cards as discussed in the Battlefield gaming demo and others. Geforce20, as stated, is not
Ray tracing ... It is
hybrid Ray tracing. They still use CUDA cores for BVH construction, they still use the rasterizer pipeline heavily... Ray trace cores are only used for intersection testing and the result looks like a grainy mess.. From there, they use tensor cores heavily to apply various forms of filtering/interpolation to produce a sensible real-world image. Like the prior implementation, this one is just a series of hacks.. Better hacks but hacks nonetheless. Hacks that fall short still of RT ray tracing. So, it's literally a prototype platform.. RT ray tracing Version 1.0.
Something underserving of an initial buy outside of developers especially at a premium.
Yeah sure, raytracing can be done on an 80286 if you want, doesn't mean it's useful for realtime. The RT cores in Turing perform very fast BVH lookups, which is where the main acceleration is gained for raytracing. Combined with limiting to small numbers of rays per pixels utilizing the Tensor cores to assist with denoising that only a vastly larger amount of rays per pixel could solve, that's 2 very specific hardware accelerated purposes that general function compute cores won't be able to match. Will AMD compute cores be faster than Pascal using DXR? Very likely and probably a significant margin. Faster than dedicated hardware solving two important issues with raytracing? Highly doubtful it will be close.
Appeal to extremes are unnecessary.
30FPS is as useless as 10FPS.
RT cores perform intersection tests
BVH is constructed in CUDA cores just like it is today.
Tensor cores help mask how crappy the end product is.
Essentially hacks atop the core functionality that is performed in CUDA cores.
Of the 2 new specific hardware accelerated portions of the cards .. One accelerates intersection testing.. The other masks how incapable the Ray Tracing methodology is. All together known as Hybrid Ray tracing. A series of hacks just like the prior version.
Interesting because this is one part software and another part hardware.
As the hardware is set upon a hacky implementation it is anyone's game as to how to exploit it in software.
As for that, it seems, when you focus moreso on the approach/software, you can get incredible results :
It was oc3d.net. It should be noted that the framerates in the graph are
MAX frames, not average. They list minimums on their page as well, but no averages. I don't see any mention of a 1080 Ti. Also interesting to note that overclocking did nothing for performance on the 2080 Ti and is possibly within normal variance on the 2080 (possible power limit?).
https://www.overclock3d.net/reviews/gpu_displays/nvidia_rtx_2080_and_rtx_2080_ti_review/31
While the 2080 is still an absurd $800. It would be considered the absolute max for most and it is a turd for
real time ray tracing.
- Hybrid Real time ray tracing : A new series of hacks to try to produce results to the end user that mimic ray tracing
Barely able to achieve FPS people equate with performance thus it will be tuned down significantly to a point that wont be much different than existing cards
- DLSS : Another series of hacks to try to feign performance that doesn't exist
- Existing HW : BVH construction is a significant part of Real time ray tracing and is done in CUDA cores both on RTX cards and non RTX cards. Existing HW doesn't have RT cores to accelerate intersection tests. However, It is already clear that ray performance per pixel is limited by this new hardware such that the bulk of the visual magic is done by tensor core approximations and hacks.
A more mature architecture is one in which BVH construction is accelerated much more.
RT cores are broken out into a separate MCM
You have far less meme learning approximate image generation via tensor cores.
Beta prototype hardware at a premium.
It's as a user said.. This is really meant for developers/Quadro cards but to make the sale there they had to produce end users thus why this is in a new line of Geforce cards. Also, due to mining setting a new price precedence and the clearly vocal group of people who will buy their wares no matter the cost, Nvidia jacked the price. Lastly, they didn't want to compete against the huge inventory of Pascal they had to sell. So, win/win across the board.
Really is a dead conversation from this point.
The compute performance is interesting but not at these prices and not due to having to wade into a whole new paradigm of
beta hardware to find potential tweaks especially with 7nm right around the corner and the power utilization off the charts. For gaming, the 1% or less will buy this no matter what.
Facebook
Twitter