NVidia Ada Speculation, Rumours and Discussion

[DavidGraham]

The fact that NVIDIA makes the numbers so nebulous is extremely annoying. SHOW us the performance WITHOUT DLSS. Then show us DLSS 2.0 vs 2.0, then 2.0 vs 3.0.

People need to chill down, the 4090 enjoys double the fp32 of the 3090 (cores + clocks), has triple the transistor budget, has better IPC for each core (Shader Execution Reordering "SER" which supposedly makes raster operations faster, providing up to 25% better gaming performance), has more capable RT cores and Tensor cores, there is no reason it can't deliver solid 2X in raster operations.

We already have the leaked Time Spy Extreme result which shows more than 80% improvements.

 

[troyan]

NVIDIA details AD102 GPU, up to 18432 CUDA cores, 76.3B transistors and 608 mm² - VideoCardz.com

NVIDIA AD102 GPU has 76.3B transistors The most transistor-packed consumer GPU yet. Not everything was revealed by NVIDIA today, but luckily for us, some board partners were quick to share the missing details. Today, NVIDIA introduced its Ada Lovelace architecture for next-gen gaming...

videocardz.com

NVIDIA RTX 6000 Ada Generation

The NVIDIA RTX™ 6000 Ada Generation combines the latest-generation RT Cores, Tensor Cores, and CUDA® cores with 48GB of graphics memory to deliver the ultimate desktop performance designers, engineers, and artists need to propel innovation forward.

www.pny.com

 

[ninelven]

With that pricing AMD doesn't really need to worry about the perf. Just launch midrange and budget low end that shreks the 3000 series. Take their market share, drink Nvidia's milkshake, the top end will take care of itself.

 

[Below2D]

Strategy by NVidia to conceal the real performance of Ada from AMD? All the marketing is in place, gamers are primed, AMD doesn't know what it's aiming at...

And gamers don't know either what they're getting.

 

[DavidGraham]

https://twitter.com/x/status/1572300538221305858

 

[Bondrewd]

With that pricing AMD doesn't really need to worry about the perf.

They still wanna be on top or somewhere there.

 

[davis.anthony]

You can't grumble at Nvidia for the software side of things, they seem to have spent some of that overpriced GPU money on new software innovations that seem fresh, useful and exciting.

Now if they can use can use RTX Remix on Crysis (2007) that would be great.

 

[hurleybird]

People need to chill down, the 4090 enjoys double the fp32 of the 3090 (cores + clocks), has triple the transistor budget, has better IPC for each core (Shader Execution Reordering "SER" which supposedly makes raster operations faster, providing up to 25% better gaming performance), has more capable RT cores and Tensor cores, there is no reason it can't deliver solid 2X in raster operations.

We already have the leaked Time Spy Extreme result which shows more than 80% improvements.

https://images.nvidia.com/aem-dam/Solutions/geforce/ada/news/rtx-40-series-graphics-cards-announcements/geforce-rtx-40-series-gaming-performance.png

First three benchmarks are raytraced games without DLSS 3, at 4K max settings, and the 4090 is averaging perhaps 1.6x performance. RT isn't exactly cheap at those settings, and the 40 series substantially improves RT performance, so take away RT and you are maybe left with a 1.4x increase in pure raster. Then consider that, more than likely, much more than likely actually, Nvidia is picking titles that make them look good.

Fury X looked like it might have be twice the performance of a 290X on paper. In reality, it was more like 1.3x. Architectural brick walls exist, and it seems Nvidia ran into one.

 

[DavidGraham]

First three benchmarks are raytraced games without DLSS 3, at 4K max settings

Are you drunk? Neither Valhalla nor Divison 2 have any kind of ray tracing or DLSS, and Resident Evil 8 has one of the more very limited and superficial ray tracing implementations in gaming.

I don't really know what you are blabbering on about.

 

[trinibwoy]

You can still use the slider on live youtubes, you know.

I know but I expected the big reveal to be at the end. Doh!

 

[hurleybird]

Are you drunk? Neither Valhalla nor Divison 2 have any kind of ray tracing or DLSS, and Resident Evil 8 has one of the more very limited and superficial ray tracing implementations in gaming.

Not drunk, just betrayed by Google where a search for "Valhalla/Division 2 Raytracing" reveals a ton of results that make them look like they support raytracing, but in each and every result "reshade" is hidden behind the ellipsis

 

[PSman1700]

Didnt expect such huge improvements in normal raster performance. Did expect it for RT/ML acceleration though

 

[troyan]

https://twitter.com/x/status/1572300538221305858

So, we can put Raytracing in thousend of games and compare RDNA3 with Lovelace? Man, that will be fun.

 

[DavidGraham]

just betrayed by Google

I can relate to that my man!

At any rate, all three of those games are AMD sponspred titles, where AMD GPUs excel at compared to their counter NVIDIA solutions, so it's natural that they score the lowest uplifts in them, this could be a deliberate choice by NVIDIA: pick examples of the worst performing titles on their hardware to show the gains in.

On another note though, here is the TSE scores if the announced cards.

https://twitter.com/x/status/1572278163199053824

 

[DegustatoR]

So I was looking for some actual performance numbers (as in opposed to words like "slow" and "embarassing").

1. CP2077 new RT Overdrive DLSS Off vs 2 vs 3 with frame generation (3FG below): 20/60/90 fps. 3X with DLSS2, only +50% with frame generation on top.
2. MSFS DLSS Off vs 3FG: 50/110, 65/130, 55/110. So roughly 2X and probably all from frame generation as the game is notoriously CPU limited.
3. A bunch of e-sport games shown here in 1440p High preset:
   1. Apex Legends at 300 fps frame cap. The game runs at 191 fps on a 3080 12GB.
   2. Fortnite at 600 fps. Couldn't find any 1440p High results. 128 fps on Epic on a 3080 12GB.
   3. Rainbow Six Siege at 500 fps. Again no High preset results. ~230 fps on Ultra on a 3090Ti.
4. This fabled masterpiece to which every expert has already triggered with words mentioned above:
   1. Resident Evil Village as 100% on a 3090Ti in 4K "highest" (presumably this means with RT). 100 fps on a 3090Ti. This puts 4080-12 at 90 fps, 4080-16 at 120 fps, 4090 at 170 fps. (3080-10 is at 75 fps there btw.)
   2. AC Valhalla as 100% on a 3090Ti in 4K "highest". 64 / 68 / 69 fps. This puts 4080-12 at 60 fps, 4080-16 at 80 fps, 4090 at 100 fps. (3080-10 is at ~55 fps here.)
   3. The Division 2 as 100% on a 3090Ti in 4K "highest". 77 fps on a 3090 here which means that 3090Ti should be 80'ish. 4080-12 at 80, 4080-16 at 95, 4090 at 135 fps.

All in all this is such a weird choice of what to show performance in that I'm with @Jawed in thinking that Nvidia is playing smoke and mirrors for now.
It's either really old games, games running at framerate caps or games with known performance issues on Nvidia h/w which may not scale well on Lovelace in the first place.
They haven't really shown much of RT performance and didn't show any modern non-RT game like Elden Ring for example.
There's also a distinct lack of Unreal Engine titles.

I definitely wouldn't make any far reaching conclusions about Lovelace performance based on this data.

 

[trinibwoy]

Nvidia sandbagging? Never thought I would see the day.

 

[hurleybird]

I can relate to that my man!

At any rate, all three of those games are AMD sponspred titles, where AMD GPUs excel at compared to their counter NVIDIA solutions, so it's natural that they score the lowest uplifts in them, this could be a deliberate choice by NVIDIA: pick examples of the worst performing titles on their hardware to show the gains in.

I wouldn't read too much into it. If anything, I would say the reverse is actually more likely. Games that are very well optimized for Ampere are not necessarily also optimized for Ada, and at the least I would not expect them to be better optimized for Ada than Ampere. Rather, new generations tend to shore up weaknesses in previous generations, so if you want the best gen-on-gen improvement you want to find titles that exploit those weaknesses. The scenarios where RDNA 3 sees the largest generational will probably be ones with heavy RT that cause RDNA 2 to falter. It's not unheard of for a sponsored game in a prior generation to outright switch preferences in a later generation.

Even though on average I take the opposite position from you, the error bars so huge that I just wouldn't read anything into a title being sponsored by either company for these comparisons.

 

[yuri]

Duh, the Turing mood again? It doesn't matter, gaming people simply buy GeForce (9:1). Is x080 unexpectedly expensive? Buy a x070 or lower, why not?

New shiny SW features are here. RT lead is sealed. GeForce brand lives on. These will sell plenty.

 

[hurleybird]

RT lead is sealed.

Is it?

 

[DavidGraham]

you want the best gen-on-gen improvement you want to find titles that exploit those weaknesses.

You have the TSE scores to guide you , I think they are enough for now.

On another note, Shader execution reordering = Coherency sorting, also apparently BVH build has been greatly accelerated, you could make an argument for Level 5 RT with this?

https://twitter.com/x/status/1572250505149136897