NVidia Ada Speculation, Rumours and Discussion

[DegustatoR]

Also people should keep in mind games have to be developed with the lowest common denominator in mind, the Series S. Based from my findings, the Series S runs texture settings that are similar to what a 4 GB PC GPU offers. That doesn't apply to PS5 ports of course.

So very likely, if you adjust texture quality settings accordingly, 8 GB and 6 GB will be enough for the whole generation.

Yes but on PC you're very likely to get a higher quality level than on XSS which will also include RT or more advanced RT.

8GBs will be enough for console level graphics for the whole duration of this generation most likely. But PCs are not about console graphics levels.

 

[PSman1700]

Well, I look forward to how Avatar game will look on 8 GB. It also supposedly will be built upon ray tracing. If it produces okay textures, than I'm okay. I'm simply too traumatized by lastgen games which had super low res, ugly textures anything below their maximum texture setting, like RDR 2/Cyberpunk/AC Origins. Some games are exception to the rule, like Doom Eternal, but sadly, an exception. To me lastgen always felt like even turning texture quality one notch always degraded and downgraded textures in a huge manner to a point where I simply start thinking there exists no texture quality.

True, but think that these games have to run on PS5 and Xbox too, this 8GB will sure be enough.

 

[arandomguy]

The max settings testing methodology while debatable from a direct user experience perspective does ensure at least an objective criteria when testing from a purely academic stand point. I think we have to acknowledge that individual game settings cannot be assumed to scale uniformly depending the hardware.

For instance also take VRAM, since that is another discussion thread here, we know VRAM amount does not uniformly impact all game settings the same way as well. Should say max texture settings no longer accommodate VRAM of a certain amount do we then separate "max textures" into it's own grouping separate from "max settings?"

I feel that if we go with RT as being separate from "max settings" we should only do so by also acknowledging that there is an inherent subjective bias with that decision. No different than adjusting any other setting or say HardOCP's testing criteria (as that was brought up). This isn't to say whether or not it is the right or wrong approach, that in itself is separate debate.

As a further addendum the prevalence vendor specific upscaling techniques that are not 1:1 substitutes of each other adds another wrinkle to the issue of actually comparing across hardware. There might end up being a large disconnect between actual usage going forward and how reviews are done if this isn't accounted for either.

I feel that it might be interesting in terms of examining how maybe hardware testing and reviews might possibly need to evolve going forward. At least with GPU testing the last major change was maybe the the growth in the look at/importance of frame times? That started around the early 2010's but otherwise has been generally stagnant otherwise. Well there's been a move more so towards youtube/video as the content platform versus written tests but that's on the presentation side.

 

[Below2D]

Also people should keep in mind games have to be developed with the lowest common denominator in mind, the Series S. Based from my findings, the Series S runs texture settings that are similar to what a 4 GB PC GPU offers. That doesn't apply to PS5 ports of course.

So very likely, if you adjust texture quality settings accordingly, 8 GB and 6 GB will be enough for the whole generation.

It will but who buys a $500 GPU to play with Medium textures?

 

[Clukos]

Edit: Nevermind, saw we are discussing the EVGA situation in another thread.

 

[Flappy Pannus]

Appropriate thread is here.

 

[Frenetic Pony]

It will but who buys a $500 GPU to play with Medium textures?

No one, doubt there'll be almost any GPUs this gen with 4gb. AMD is moving to Apple's model of big integrated GPUs, at least on the low end so they'll not be putting anything out with less than 8gb for dedicated again.

Not sure what Nvidia will do though. They're kind of stuck on the lower end, AMD laptops don't need Nvidia low end dedicated hardware, and eventually Intel won't either. The M series will go bye bye, though I can see a low end dedicated card with 6gb for Nvidia. GDDR is expensive, they could knock a good $50-60 off the 3060 by getting rid of 6gb of ram. Who knows, maybe there'll be a 4050 with 6gb for $199. Something "good enough" for the person that wants to play Dota at 100fps and/or encode higher quality outputs from their console.

 

[Silent_Buddha]

The max settings testing methodology while debatable from a direct user experience perspective does ensure at least an objective criteria when testing from a purely academic stand point. I think we have to acknowledge that individual game settings cannot be assumed to scale uniformly depending the hardware.

For instance also take VRAM, since that is another discussion thread here, we know VRAM amount does not uniformly impact all game settings the same way as well. Should say max texture settings no longer accommodate VRAM of a certain amount do we then separate "max textures" into it's own grouping separate from "max settings?"

I feel that if we go with RT as being separate from "max settings" we should only do so by also acknowledging that there is an inherent subjective bias with that decision. No different than adjusting any other setting or say HardOCP's testing criteria (as that was brought up). This isn't to say whether or not it is the right or wrong approach, that in itself is separate debate.

As a further addendum the prevalence vendor specific upscaling techniques that are not 1:1 substitutes of each other adds another wrinkle to the issue of actually comparing across hardware. There might end up being a large disconnect between actual usage going forward and how reviews are done if this isn't accounted for either.

I feel that it might be interesting in terms of examining how maybe hardware testing and reviews might possibly need to evolve going forward. At least with GPU testing the last major change was maybe the the growth in the look at/importance of frame times? That started around the early 2010's but otherwise has been generally stagnant otherwise. Well there's been a move more so towards youtube/video as the content platform versus written tests but that's on the presentation side.

Sure testing with everything maxed including RT is certainly a technically interesting, even if not terribly useful metric for many users and as such should be included in every review.

If all the review is going for is technical capabilities then that's fine.

If a review is also attempting to present graphics cards as realistic gaming experiences then RT is required just as are high, medium and even low settings. Most sites will obmit low and some only go down as far as high.

Similar to how 4k, 1440p and 1080p are still generally tested by sites that attempt to provide useable benchmark metrics for the average PC gamer.

If all that mattered was the purely technical aspect of what was objectively the most technically capable card, then I would fail to see why anything less than 4k native resolution with RT would ever need to be tested.

At some point most review sites want to at least attempt to give their readers an idea of how a graphics card might potentially perform for them while gaming at an acceptable framerate at their chosen resolution. And for many people that means RT off.

I only ever look at review sites who also provide high and medium graphical settings in addition to ultra presets. It gives me a general idea of where I'll have to dial in settings and makes it easier to determine which settings I might want to disable to gain the performance I need. Bonus points for any site that might also provide low graphics presets.

Regards,
SB

 

[DegustatoR]

Lenovo GeForce RTX 4090 GPU pictured some more, it is 36 cm long - VideoCardz.com

Lenovo RTX 4090 pictured by Lenovo manager Wolfstame, Lenovo China Gaming Desktop Product Planning Manager, posted new photos of the upcoming RTX 4090 GPU. The new GPU has been pictured alongside the Lenovo gaming laptop, it is massive, that’s for sure. The Legion Y9000X is an 16-inch laptop...

videocardz.com

And for many people that means RT off.

I wonder how many people would go for 4090 and play with RT off... Me personally hasn't been opting for such even on a 2080. Framerates were fine.

 

[trinibwoy]

GPU cooling solutions are bordering on ridiculous. How much bigger and heavier can these things get?

 

[nutball]

GPU cooling solutions are bordering on ridiculous. How much bigger and heavier can these things get?

If they want me to buy one, no bigger at all. I don't just mean the 4090, I mean all of it. It's silly flash-backs and PTSD from the bad old days.

Which does make me wonder...

 

[DavidGraham]

Before the announcement, I am going to go out on a limb here and claim that if Ada is truly a 75 billion chip compared to 28 billion of Ampere, then Ada is going to be significantly faster than 2X, maybe 2.5X or more, that ... or Ada has another focus entirely.

Let's trace it back a little. the G80 (8800 GTX) had 680 million transistors, more than double that of the 7900GTX (278 million), it successfully achieved almost double the performance. Tesla (GTX 285) doubled the transistor count to 1400 million (1.4 billion) too, but it had a compute focus, so the transistor budget went there and the uplift in performance was limited to around 50%, Fermi (GTX 580) doubled that again to 3 billion (again with a heavy compute focus), so got around 60% more performance.

The pattern quickly got corrected with Kepler (GTX 780Ti), which more than doubled the budget to 7 billion, while successfully almost doubling performance. Maxwell (Titan X) continued down the path and achieved 30% more performance with a mere 14% increase in budget to 8 billion. Pascal (Titan Xp) saw a modest budget increase to 12 billion (a 50% increase), while also achieving almost double the performance, Turing (Titan RTX) boosted the budget to 18 billion (another 50%) achieving 40% more performance, but it had a different focus (Ray Tracing and Machine Learning), so by those metrics it often achieved triple the performance for those metrics. Ampere (3090Ti) continued down that path, increasing budget by 50% to 28 billion, to achieve 50% more performance.

Data Center GPUs have forked themselves away from consumer GPUs with the V100/A100/H100 lineup, so we won't see the budget wasted on compute again, and with Ray Tracing and Machine Learning already paid for, If Ada is truly going from 28 billion to 75 billion, a 2.7X increase in budget (this explains the huge uplift in power), it will either net us a more than 2X the performance, or the focus this time is on something else entirely, something mysterious.

 

[Qesa]

There's the big L2$ to consider. Nvidia is paying a lot of transistors to increase performance while barely budging on memory bandwidth. How would your analysis go if nvidia opted for HBM instead of a large cache to get the same performance with far fewer transistors?

Though with that said, 75B still seems unreasonable for a ~600mm^2 chip. And didn't kimi also vastly overestimate GH100's count before launch? I'd personally expect it to land around 60B

 

[SlmDnk]

The countdown looks like a real-time animation... EDIT: Downloaded the video, which is 240KB in size, so it's just a 1-second/30fps video looped back and forth with the clock screen superimposed. That's why the right side screen is not affected by the pulsating light.

 

[DavidGraham]

There's the big L2$ to consider. Nvidia is paying a lot of transistors to increase performance while barely budging on memory bandwidth. How would your analysis go if nvidia opted for HBM instead of a large cache to get the same performance with far fewer transistors?

Though with that said, 75B still seems unreasonable for a ~600mm^2 chip. And didn't kimi also vastly overestimate GH100's count before launch? I'd personally expect it to land around 60B

Kimi overestimated Hopper's die size, not transisor count, and NVIDIA hinted at that number already in their teasers.

The numbers shown on a sticky note may refer to the specs of the AD102 GPU. This could be the size of the die (629 mm²) and the number of transistors (75.38B).

NVIDIA confirms upcoming Ada architecture - VideoCardz.com

NVIDIA shares a teaser for the upcoming Ada architecture NVIDIA released a new teaser for its “GeForce Beyond” broadcast next week. It has now confirmed the name upcoming architecture. NVIDIA teases Ada Architecture, Source: NVIDIA The Twitter teaser posted a “Diagram for the computation by the...

videocardz.com

 

[gamervivek]

2x would've been possible even if it were just a scaled-up Ampere since from the very first rumors it was a proper scaling up and not just +1 GPC like 3090 had over 2080Ti. Combine that with 30-40% higher clocks than what 30xx hobbled to on Samsung's process and 2x was always on the cards.

How much higher nvidia can go or want to go is up for speculation, alongwith the competition's next-gen performance. Pascal could've been a bigger increase over Maxwell if nvidia bothered with a 600mm2 chip.

 

[Jawed]

Maybe ray sorting will eat up a lot of transistors?

 

[Kaotik]

The transistor count + die size theory has just one big issue: H100 is 80bn @ 814mm^2 on TSMC N4 (or was it 4N, whichever). 75bn @ 629mm^2 on same (or very similar) process doesn't add up.

 

[DegustatoR]

The transistor count + die size theory has just one big issue: H100 is 80bn @ 814mm^2 on TSMC N4 (or was it 4N, whichever). 75bn @ 629mm^2 on same (or very similar) process doesn't add up.

More cache?

 

[Kaotik]

More cache?

Over 21 % better density from just more cache?