-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
NVIDIA ‘Ampere’ 8nm Graphics Cards
- Thread starter LoadsaMoney
- Start date
Soldato
I agree with kapp he seems to no his tech good times ahead but prices will be insane and I wouldn’t be a early adopter this time after the my rma with my 2080ti.
Last edited:
Soldato
I think kaap works for nvidia lol.
Soldato
What if I told you that AMD RT will be more performant than SLI Titan RTX?
Also second gen Nvidia RT is a significant upgrade.
But even then they will still be a case of optimising what uses RT, just nowhere as problematic as Turing. The 20 series will depreciate very fast for RT performance.
Soldato
Fairly hopeful for decent perf increases across the range of cards, separate from RT but there should be good increases there too. Since they released Super cards it makes me think the next cards will be even better, but hopefully those Super types of cards don't become a new milking trend instead.
I have always been wondering why only 68 RT cores in RTX 2080 Ti. What if they increase them to 512 in the second-gen RTX product.......
Soldato
- Joined
- 19 Oct 2008
- Posts
- 6,003
2H was what I was expecting, Q2.
Nice to see more competition from AMD but I think NV will take another leap ahead. It'll be 7nm done properly and much improved RT too although RT is going to take leaps forward for years I think.
Nice to see more competition from AMD but I think NV will take another leap ahead. It'll be 7nm done properly and much improved RT too although RT is going to take leaps forward for years I think.
Soldato
From what i know, Nvidia shouldn't have released Turing as RT parts to consumers. If anything they were likely going to be Quadro only to begin with considering the non RT turing line.
Do I not recall tests showing Quake RTX running at 4k 60 fps with two 2080 Tis?
RTX on Turing was always a taster.
Soldato
I'm hoping AMD smashes Nvidia in the same way they are doing with Intel (I think Intel are still in denial with their current crop of 10nm chips even with their massively reduced pricing), hopefully this will bring down prices to a more "normal" level" than the stratospheric prices we are currently seeing, Intel coming into this mix in 2020 also gives me some home for the GPU market in 2020.
Soldato
ray tracing Quake used a path tracer before they added RTX to it. RT hardware just helped accelerate the performance.
But in general, Path tracing is more physically realistic and heavier than basic RT. The only reason it works so well on quake RT is due to how simple the geometry is.
Man of Honour
I would not take those claims seriously.
I would not take those claims seriously.
I would not take those claims seriously.
Never never never take anything like this seriously until you have the hardware in your hands and properly tested.
Ray Tracing is going to continue to be a hardware killer from both NVidia and AMD for the next 3 to 5 years despite whatever claims they would like to make to the contrary.
The RTX implementation used in Quake 2 does not use geometry, etc. optimisations like Minecraft neither does that help it in terms of performance - the backend renderer will work within reason with any scene complexity within the constrains that are used to stop it going off and sucking up crazy amounts of processing time and scene complexity (within reason) only has a minor impact on its performance. Due to the nature of how it works the impact of polygon count on path tracing tends to be in the region of 1-3% of its performance load in a video game setting.
I've spent a lot of time messing with the Quake 2 RTX renderer and most of what people say about it is BS - while I'm limited in terms of showing off its potential by the geometry limits in Quake 2 - small detail tends to just break or not be possible to do with minimum geometry size and the provided materials in Quake 2 RTX lacks a wider range suited to environments outside of the original game setting that isn't the limits of the renderer used - quick collage of some of the stuff I've been messing with:
The main problem with the renderer used in Quake 2 really is that it needs decent framerates and the hardware to do more rays per pixel than is currently possible so as to minimise artefacts from the temporal way it samples the world resulting in noise and sometimes smearing of foreground objects over background ones - both of which are possible to mitigate to a level most people would no longer notice them within a couple of hardware generations tops.
EDIT: For reference my GTX1070 renders these scenes at 800x600@30FPS, 1280x720@18FPS - anything above that the framerate counter breaks and it just shows 10FPS - in reality it is doing around 7FPS in those scenes at the original resolution (1440p) I took the screenshots in. Turing cards are approx. 6x faster.
This isn't like current games using RTX/RT that just use it for one lighting component - the implementation in Quake 2 is doing all lighting, GI, caustics and reflections using path tracing - everything in the scene is being done via RTX no elements of the scene are relying on old techniques.
Last edited:
Soldato
Oh i heard it all on good grounds from more reliable sources than Adored has.
Soldato
- Joined
- 21 Apr 2007
- Posts
- 2,544
That's exactly what is needed though as Pascal > Turing was so naff there is lost ground to make up for. Anything RT related is a bonus but not the main course, its easy to get distracted by the marketing and it gonna be smoke and mirrors all the way up to the release of Cyberpunk.
If Nvidia keeps RT, they either need to find a more efficient way of doing it or otherwise they'll have to stick more RT cores onto the die. The current 2080ti has 10% of it's die taken up by RT cores, that's 77mm2 of die space just for RT cores on 16/12nm.
Assuming the news is correct and Samsung has better yields than TSMC - perhaps they can produce bigger dies - let's say 400mm2 from Samsung. So you can increase the RT space from 77mm2 to 100mm2, leaving 300mm2 for your conventional raster architecture. So assuming a linear scale - you have 50% greater density and an extra 30% space for the RT cores. Maybe they find a little bit of efficiency somewhere and what you end up with is 100% extra ray tracing performance on 7nm.
And yet it's still not good enough for 4k 60fps at Ultra settings unfortunately, without using tricks like DLSS etc.
Assuming the news is correct and Samsung has better yields than TSMC - perhaps they can produce bigger dies - let's say 400mm2 from Samsung. So you can increase the RT space from 77mm2 to 100mm2, leaving 300mm2 for your conventional raster architecture. So assuming a linear scale - you have 50% greater density and an extra 30% space for the RT cores. Maybe they find a little bit of efficiency somewhere and what you end up with is 100% extra ray tracing performance on 7nm.
And yet it's still not good enough for 4k 60fps at Ultra settings unfortunately, without using tricks like DLSS etc.
Last edited:
Surely there has to be more to it than that also though, like improving efficiency of each core etc. Like with rasterising, it is not always about how many cores there are as efficiency also increases per core also?
Remember what they did with physx cards, I would rather they left the 3080 series alone so they can better play 4k and find a way to make RT work like an addon card.
It works for me i get more speed, I will not moan and if i want Raytracing could i not have the entire pcb for RT cores? If it takes up 10% what happens if i take up 100% for RT cores? I know addon cards are unpopular but it sometimes works.
It works for me i get more speed, I will not moan and if i want Raytracing could i not have the entire pcb for RT cores? If it takes up 10% what happens if i take up 100% for RT cores? I know addon cards are unpopular but it sometimes works.
Last edited:
