[HardwareLeaks/_rogame] - Exclusive first look at Nvidia’s Ampere Gaming performance

[pas008]

10 minutes ago, Sampsy said:

If that's the 3090 those numbers really suck - I don't get why people are impressed. The 2xxx series was really underwhelming as well and it came out nearly two years ago If all they can do in that time is a measly +30% performance it better be cheap.

not when you count in raytracing by far

you might think its a gimmick but its the future, devs wanted it for decades

[porina]

It does feel like we're on the edge of a new way of thinking about GPU performance. It could well be they don't significantly increase non-RT performance in absolute terms, and use updated DLSS to give more of a boost, as well as in parallel throwing better RT into the mix. We're going to need separate measures on each of those areas, both against previous gen and vs AMD (and eventually Intel too).

[D13H4RD]

30% seems just about normal-ish for a GPU performance bump generation-by-generation 

[mr moose]

1 hour ago, Sampsy said:

 If all they can do in that time is a measly +30% performance it better be cheap.

 

Not measly

 

480 to 580 was about 16%

580 to 680 was about 45%

680 to 780 was about 28%

780 to 980 was about 31%

980 to 1080 was about 62%

 

https://www.techspot.com/article/1191-nvidia-geforce-six-generations-tested/

 

Assume the two outliers are not representative of the whole (good statistical practice when working with averages especially when they are so far from the norm) and you have an average of 35% improvement.    So 30% is well within normal improvement range for a new product.

 

Or if you want to include the 980 ti because it was a year in between the 980 and 1080

980 to 980ti was 29% and 980ti to 1080 was 26.5%  bringing the release average to 31.8% or 32% if we ignore the outliers.

[D13H4RD]

Regarding this entire thing

 

1. It's a rumor
2. It's a rumor
3. It's a rumor

Did I say it's a rumor? Yeah, well, it's a rumor.

 

Best to keep expectations in check because we barely know anything else about the card that's even moderately confirmed. A 30%-ish performance improvement is well within range for NVIDIA's typical performance improvements generation-over-generation, but that also depends on the price. Turing itself was a solid 20-30% improvement over Pascal in terms of performance but the RTX price premium ended up making them worse value outright, especially at launch.

 

If Ampere's raw power proves to be a little underwhelming, I'm kinda betting that NVIDIA may not be focusing as much on raw compute power and focusing more on artificial intelligence and deep learning with stuff like DLSS.

[Loote]

1 hour ago, mr moose said:

480 to 580 was about 16%

580 to 680 was about 45%

680 to 780 was about 28%

780 to 980 was about 31%

980 to 1080 was about 62%

480->580 40nm->40nm = 16%
580->680 40nm->28nm = 45%
680->780 28nm->28nm = 28%
780->980 28nm->28nm = 31%
980->1080 28nm-> 14nm = 62%

 

Generational improvements alone are on average less than 30%, however coupled with node change you can easily expect >50%(if it was so simple)

[mr moose]

1 minute ago, Loote said:

480->580 40nm->40nm = 16%
580->680 40nm->28nm = 45%
680->780 28nm->28nm = 28%
780->980 28nm->28nm = 31%
980->1080 28nm-> 14nm = 62%

 

Generational improvements alone are on average less than 30%, however coupled with node change you can easily expect >50%(if it was so simple)

 

980ti to 1080 was 30% on jump to 16nm node.

and

1080 to 2080 was also a node jump (although slightly smaller) and 21%.  But you are right in that there is a lot more to it than that, there are increases in # cores and memory improvements and so on.    I think for the purpose of what to expect 30% is not unreasonable or measly based on being historically evident even between node jumps.

 

[AdmiralMeowmix]

The numbers here do seem weird to me considering previous leaks. These numebrs just seem too low in comparison of what other leakers have talked about in regards to performance and the changes made to get performance. But everything is a grain of salt so who knows hot this'll turn out.

[xAcid9]

Meh if this is the top end chip. But if RT performance drop is below 30 or 20, i'm sold. 

[RejZoR]

2 hours ago, mr moose said:

 

980ti to 1080 was 30% on jump to 16nm node.

and

1080 to 2080 was also a node jump (although slightly smaller) and 21%.  But you are right in that there is a lot more to it than that, there are increases in # cores and memory improvements and so on.    I think for the purpose of what to expect 30% is not unreasonable or measly based on being historically evident even between node jumps.

 

It's also important to note GTX 1080Ti made massive jump in performance within same series. A rather unusually large jump actually. So large that going to RTX 2080 or even RTX 2080Ti doesn't make sense unless you have money to piss or you specifically want RTX functionality.

[uzzi38]

11 hours ago, Loote said:

I believe just the node shrink should allow them to pack twice as many transistors, but it's more complicated than that and it could be occupied by even more RT cores.

 

 

This doesn't  seem right. I'm getting frustrated by leaks, I wish we'd know what's coming already.

As for point one yes, but that's only if you keep die sizes the same.

 

I highly doubt that Nvidia will be selling consumers 700+mm^2 dies yet again with available N7 supply and price per wafer. Not only do they not have enough supply to meet market demands, they also wouldn't have much margins at all on these dies.

 

As for architectural improvements, really? It doesn't feel right? Come on man, what I posted isn't even a leak it's a guy's talking about CUDA11 which has been publically released, and that after some playing around it not much seems to be different with Ampere vs Turing.

 

It doesn't feel right isn't a counter-point.

[Loote]

1 hour ago, uzzi38 said:

As for point one yes, but that's only if you keep die sizes the same.

Sure, they often did smaller dies on a new process and then only the next generation of the same process was big.

 

1 hour ago, uzzi38 said:

It doesn't feel right isn't a counter-point.

It was more of an agreement with:
 

Quote

Consumer facing Ampere I'd hope is different, but I can't see some revolutionary jump in performance coming as a result of this.

Based on what little info there is we start to come to conclusions, but afaik there's no way to tell which GPU would that be, how big is the influence of it being an engineering sample, or is it at all legit. I am not going to use stronger words because I'm not qualified to guess and people who are qualified are most likely forbidden from telling us.

[CarlBar]

16 hours ago, Loote said:

I believe just the node shrink should allow them to pack twice as many transistors, but it's more complicated than that and it could be occupied by even more RT cores.

 

I believe NVIDIA themselves have indicated it has 4x the RT cores. From memory of a breakdown of the Turing dies they took up quite a big area, and these RT cores have additional instruction support. I suspect most of even the 3090 die area is going to be RT cores. So this sounds quite reasonable to me.

[Metal_HorseCrab]

People here thinking the gains are impressive, oh how i miss the older days.

[pas008]

1 hour ago, CarlBar said:

 

I believe NVIDIA themselves have indicated it has 4x the RT cores. From memory of a breakdown of the Turing dies they took up quite a big area, and these RT cores have additional instruction support. I suspect most of even the 3090 die area is going to be RT cores. So this sounds quite reasonable to me.

traversal coprocessor rumor?

 

https://www.tweaktown.com/news/73209/geforce-rtx-3080-and-3090-rumored-to-pack-traversal-coprocessor/index.html

 

[porina]

5 hours ago, pas008 said:

traversal coprocessor rumor?

 

https://www.tweaktown.com/news/73209/geforce-rtx-3080-and-3090-rumored-to-pack-traversal-coprocessor/index.html

 

From the link:

Quote

I'm sure that when you saw the headline, or read this news or saw a post on social media about a "traversal coprocessor" on NVIDIA's new GeForce RTX 3080 and GeForce RTX 3090 graphics cards that you thought "what the HELL is a traversal coprocessor".

Almost correct, but I used a stronger 4 letter word where the all caps go  

 

In short, the claim here is that RT would be offloaded from the main GPU die itself. Interesting stuff. You could look at it like a coprocessor. We have to assume that on-card bandwidth wont hinder this kind of operation if it goes ahead. It also implies that non-RT parts are more of a possibility as you don't need to decide to implement it at a silicon level, but can do so at a card level.

[pas008]

50 minutes ago, porina said:

From the link:

Almost correct, but I used a stronger 4 letter word where the all caps go  

 

In short, the claim here is that RT would be offloaded from the main GPU die itself. Interesting stuff. You could look at it like a coprocessor. We have to assume that on-card bandwidth wont hinder this kind of operation if it goes ahead. It also implies that non-RT parts are more of a possibility as you don't need to decide to implement it at a silicon level, but can do so at a card level.

this means easier binning also?

[jasonc_01]

Not impressive if the pricing structure stays the same or greater than the 20 series. 

 

However if your getting a 30% performance uplift over the 2080ti at 10 series pricing, then it's impressive/palatable.

 

I could also win the lottery and wifey Cara Delevingne.

[leadeater]

4 hours ago, porina said:

In short, the claim here is that RT would be offloaded from the main GPU die itself. Interesting stuff. You could look at it like a coprocessor.

PhysX cards v2.0? 

[RejZoR]

21 hours ago, leadeater said:

PhysX cards v2.0? 

In a way I'm still shocked HW accelerated physics aren't a thing yet in games as a defacto standard. And because there really isn't any standard for it, it'll just remain a pointless gimmick shilled by NVIDIA. Because no one can implement it as a gameplay mechanic unless everyone can do it. Half-Life 2 was really the only serious attempt in that direction and how many years has passed since? A lot. And no game comes even close to Half-Life physics that were actually usable within gameplay. Even if it was just tossing radiators at headcrabbed zombies. 99% of games today don't even try that.