Question 'Ampere'/Next-gen gaming uarch speculation thread

[Ottonomous]

How much is the Samsung 7nm EUV process expected to provide in terms of gains?
How will the RTX components be scaled/developed?
Any major architectural enhancements expected?
Will VRAM be bumped to 16/12/12 for the top three?
Will there be further fragmentation in the lineup? (Keeping turing at cheaper prices, while offering 'beefed up RTX' options at the top?)
Will the top card be capable of >4K60, at least 90?
Would Nvidia ever consider an HBM implementation in the gaming lineup?
Will Nvidia introduce new proprietary technologies again?

Sorry if imprudent/uncalled for, just interested in the forum member's thoughts.

 

[Glo.]

People talk always about AMD hype trains, and always fail, competely, to see Nvidia hype trains.

Just from this thread, and first pages.
1) Ampere gaming cards H1 2020,
2) Ampere Gaming cards way more efficient, than Turing,
3) Ampere Gaming Cards made on 7 nm EUV process from Samsung.

Its always AMD. Its never Nvidia fanboys, who are responsble for hype trains, of their own, beloved, brand.

 

[Gideon]

People talk always about AMD hype trains, and always fail, competely, to see Nvidia hype trains.

Just from this thread, and first pages.
1) Ampere gaming cards H1 2020,
2) Ampere Gaming cards way more efficient, than Turing,
3) Ampere Gaming Cards made on 7 nm EUV process from Samsung.

Its always AMD. Its never Nvidia fanboys, who are responsble for hype trains, of their own, beloved, brand.

I get why it is at least partially the case. Remember poor Volta and all Vega hype vs what was delivered? The last time Nvidia failed so hard was FX 5800 in 2002. GTX 280 and GTX 480 weren't great either, but at least they were the fastest thing available at release (and had no ridiculous will-stom-the-competition ad campaigns).

 

[raghu78]

People talk always about AMD hype trains, and always fail, competely, to see Nvidia hype trains.

Just from this thread, and first pages.
1) Ampere gaming cards H1 2020,
2) Ampere Gaming cards way more efficient, than Turing,
3) Ampere Gaming Cards made on 7 nm EUV process from Samsung.

Its always AMD. Its never Nvidia fanboys, who are responsble for hype trains, of their own, beloved, brand.

People have been accustomed to Nvidia winning by a big margin and having a huge lead in performance and efficiency( perf/watt and perf/sq mm). But I feel that has also led to this thought process that Nvidia can never make a mistake and that they are invincible. This strategic blunder by Nvidia to bet on Samsung is already showing up in the power specs. When we have the press reviews if these power numbers turn out true then Nvidia would have the worst perf/watt improvement gen on gen in a decade after Fermi. From what we already know about the Xbox Series X GPU and Microsoft's statement that the power draw is 1x of Xbox One GPU, RDNA2 is a very power efficient GPU architecture. Anyway the initial signs are there for everyone to take notice of but if they still want to ignore it thats upto them.

 

[GodisanAtheist]

People talk always about AMD hype trains, and always fail, competely, to see Nvidia hype trains.

Just from this thread, and first pages.
1) Ampere gaming cards H1 2020,
2) Ampere Gaming cards way more efficient, than Turing,
3) Ampere Gaming Cards made on 7 nm EUV process from Samsung.

Its always AMD. Its never Nvidia fanboys, who are responsble for hype trains, of their own, beloved, brand.

- To be fair, NV has pulled some wild rabbits out of their hats, even when the only people they were competing with were themselves. Further still, NV's successes were amplified by AMD's colossal failure to rise to the moment.

The Maxwell bump over the Kepler line on the same 28nm node was impressive, and only made to look better by AMD basically just rebadging their entire line-up *AGAIN* and putting the Fury series on top.

The Pascal Line was a huge jump over Maxwell and only made to look better by AMD hyping their own products in Polaris and Vega, then failing to get anywhere near the 1080ti.

The Turing line was the weakest launch from NV yet, but came out nearly totally unopposed by AMD with the exception of the Vega 2 which was just a rehashed but die shrunk Vega 64. The 5700XT brought the cost down but didn't move the performance needle.

With Ampere we went from "NV is deep in a hole and stuck with Samsung" to NV's line-up looks like its ready to launch before we've seen or heard anything about the RDNA2 cards sans the console announcements. It just feels too much like the Polaris/Vega hype up all over again.

I like that AMD has become more nimble in recent years with the launch of Vega 2 and the Navi line, getting something in front of customers even if its not the ideal product. That being said, AMD is unarguably starting further behind the starting line on this latest gen than they ever have before and people have been hurt way too many times.

 

[Glo.]

With Ampere we went from "NV is deep in a hole and stuck with Samsung" to NV's line-up looks like its ready to launch before we've seen or heard anything about the RDNA2 cards sans the console announcements. It just feels too much like the Polaris/Vega hype up all over again.

I like that AMD has become more nimble in recent years with the launch of Vega 2 and the Navi line, getting something in front of customers even if its not the ideal product. That being said, AMD is unarguably starting further behind the starting line on this latest gen than they ever have before and people have been hurt way too many times.

Right.

And what picture does it paint for AMD's product competitiveness the fact that RTX 3080 has to be made with 102 die, instead of typical 104?

What picture does paint the fact that Ampere GPUs will draw stupid amounts of power, because of clocks being required to be so high?

What picture for AMD GPUs does paint a fact that Nvidia is completely rushing the release of those GPUs?

Does it feel IN ANY WAY Polaris/Vega hype train all over again?

 

[Gideon]

What picture for AMD GPUs does paint a fact that Nvidia is completely rushing the release of those GPUs?

While I agree with your other points, where did you get that. Considering how ancient Turing is, it doesn't seem all that rushed. Especially when considering that they already released the server parts months ago.

Also what's your estimation for the performance uplift of Navi 21 vs 5700 XT (and 3090 vs 2080ti)?

Are you honestly expecting Big Navi to have ~225% the performance of a 5700XT (which would be required for a clear win, if Nvidia can pull off a super dissapointing 40% uplift) ... and at @275W ?

EDIT:
nvm , answered in the other thread

 

[Glo.]

While I agree with your other points, where did you get that. Considering how ancient Turing is, it doesn't seem all that rushed. Especially when considering that they already released the server parts months ago.

Also what's your estimation for the performance uplift of Navi 21 vs 5700 XT (and 3090 vs 2080ti)?

Are you honestly expecting Big Navi to have ~225% the performance of a 5700XT (which would be required for a clear win, if Nvidia can pull off a super dissapointing 40% uplift) ... and at @275W ?

EDIT:
nvm , answered in the other thread

Where am I getting that last point.

If Nvidia is not rushing those GPUs out, why is RTX 3080 on 102 die, and not 104, like it ALWAYS was, before?

 

[n0x1ous]

Where am I getting that last point.

If Nvidia is not rushing those GPUs out, why is RTX 3080 on 102 die, and not 104, like it ALWAYS was, before?

X80 was big die on Kepler. Not that abnormal

 

[IntelUser2000]

So excuse me if I don't take AMD's RDNA2 claims at a face value. They claim another 50% perf/watt increase over RDNA. Sure, RDNA did that over Vega. I'll believe it when I see it.

Perhaps they are also downplaying the gains for moving to 7nm. It's definitely a enabler. Without it, it might have been 10% better.

Of course same goes for everyone.

How much perf/watt gains Polaris and Vega was supposed to have over predecessors?

The Pascal Line was a huge jump over Maxwell and only made to look better by AMD hyping their own products in Polaris and Vega, then failing to get anywhere near the 1080ti.

Pascal was good in perf/watt and mm2 over Maxwell, yes. But that was also largely due to process and it wasn't that exciting in the perf/$ at all. GTX 1080 improved performance, but also raised prices and started the controversial FE pricing. 1080 Ti was only good because they decided to reasonably price it. Remember how badly priced the Titan version was?

Maxwell only looked good because previous generations were meh due to staying in the same process.

 

[moonbogg]

Rumors have been suggesting process nodes that go by the marketing names 8nm, 7nm, and as low as 5nm for these cards. I'm just curious if processes getting smaller and smaller means we are approaching the end of the road in terms of density. Performance improvements have still been decent, but certainly seem to be falling on a flattening curve. How long do we have before GPUs are marketed in terms of new "features" instead of performance increases much like Intel CPUs now are?

 

[jpiniero]

Rumors have been suggesting process nodes that go by the marketing names 8nm, 7nm, and as low as 5nm for these cards. I'm just curious if processes getting smaller and smaller means we are approaching the end of the road in terms of density. Performance improvements have still been decent, but certainly seem to be falling on a flattening curve. How long do we have before GPUs are marketed in terms of new "features" instead of performance increases much like Intel CPUs now are?

There's always room to go bigger, that is if there's people willing to pay...

 

[moonbogg]

There's always room to go bigger, that is if there's people willing to pay...

Based on my passive and very lamen observations, my gut tells me that in 5 to 10 years something fundamental will have to change because process improvements will just become impractical. After taking improvements for granted and getting to experience the golden days of hardware improvements, it's hard to imagine watching technology reach "maturity" where it can't reasonably be improved much further at a reasonable cost.

 

[IntelUser2000]

I'm just curious if processes getting smaller and smaller means we are approaching the end of the road in terms of density.

Here are the basic dimensions of TSMC's 7nm process:
Pitch 30 nm
Width 6 nm
Height 52 nm
Contacted Gate Pitch (CPP) 64 nm (HP)
57 nm (HD)
Minimum Metal Pitch (MMP) 40 nm

The dimensions decrease approximately by 70% if the density gains are 2x. In the post-3nm TSMC timeframe, it'll be too small to reliably/cheaply produce so things will change.

Hot Chips 2020 Live Blog: Intel's Raja Koduri Keynote (2:00pm PT)

www.anandtech.com

Starting from 05:20PM EDT it shows the direction we're headed.
-Pitch Scaling: It shows 3x gains from Intel's 10nm process. 7nm should be 2x that. TSMC's 3nm is up to 35% better in density so they are pretty much at limits. Some think Intel's 7nm is greater than 2x in density over the predecessor so even for Intel they are either at limits or close to by 7nm timeframe.
-Nanowire: By putting more nanowires we increase gate density without scaling pitch
-Stacked Nanowire: Stacking the nanowires mean another 2x
-Wafer to Wafer stacking
-Die to Wafer stacking

Wafer to Wafer and Die to Wafer will increase costs, but there are other advantages that will come with it and some markets can absorb the costs. Also, unlike Flash NAND, stacking won't be more than 2 layers. Really Intel's 7nm and TSMC's 3nm are where traditional scaling based on pitch reduction ends.

it's hard to imagine watching technology reach "maturity" where it can't reasonably be improved much further at a reasonable cost.

Same thing happened with transportation. In the 70's some were expecting FTL speeds and travelling to Pluto by 2000.

I have a Popular Science article where they were expecting supersonic travel by 2015. They gave up because they were aiming for existing subsonic airplanes in terms of efficiency and pollution generated, not something that'll be available in future subsonic airplanes.

The same guys expecting linear progression(or in other words, just plotting line straight into the future) are those that expect general AI by 2050s.

 

[GodisanAtheist]

Right.

And what picture does it paint for AMD's product competitiveness the fact that RTX 3080 has to be made with 102 die, instead of typical 104?

- Nothing, it doesn't tell us anything about AMD's competitiveness. Maybe NV isn't happy about the lift over Turing,maybe they're looking to leverage the shrink into as large a jump as they can make, maybe they smell blood in the water and want to leave AMD in the low margin sector only. We don't know.

What picture does paint the fact that Ampere GPUs will draw stupid amounts of power, because of clocks being required to be so high?

- Again, a bunch of speculation that is, as far as I can tell, based on a connector that can theoretically carry oodles of power...but doesn't have to. If the idea is to implement a new GPU power connector standard, they can use it on cards that use 300 or 250 or 150 or whatever watts of power just to leverage their market position.

Additionally, the market has shown that it will accept high power draw if the benefit is halo performance (AKA Fermi). What the market won't tolerate is higher power draw for worse performance.

What picture for AMD GPUs does paint a fact that Nvidia is completely rushing the release of those GPUs?

- Nothing about NV's launch timing is rushed. They're following their standard cadence. What does it tell you that AMD is behind schedule? Nothing, because we don't know that they are and even if we did it wouldn't mean anything concrete.

Does it feel IN ANY WAY Polaris/Vega hype train all over again?

-High expectations for the AMD product to match or overtake the equivalent NV products in price/performance/efficiency and the associated nervous enthusiasm of the folks that profess as much.

 

[Veradun]

I get why it is at least partially the case. Remember poor Volta and all Vega hype vs what was delivered? The last time Nvidia failed so hard was FX 5800 in 2002. GTX 280 and GTX 480 weren't great either, but at least they were the fastest thing available at release (and had no ridiculous will-stom-the-competition ad campaigns).

I'm ready to bet Vega sold much more gaming cards than Volta :>

 

[Gideon]

So just to recap:

We know now that 3090 is indeed the rumored high-TBP (~350W) monster. This also means that it's almost certainly build on Samsung 8nm (not that I doubted that, but some here still did).

What people get wrong IMO, is expecting it to perform badly because of the chosen high power consumption. I wouldn't bet on that, until we get at least some evidence to support it.

According to Igor's lab The high TBP is not only because of 8nm but also because of really power-hungry GDDR6X. (Igor estimated 230W for GPU only).

Nvidia might have needed to clocked the flagship out of it's comfort zone simply to reach the performance target they initially had (for 7nm and a more power-efficient GDDR6X). This in turn was probably necessary to get people to upgrade.

GDDR6X might have missed it's design spec slightly (just as initial HBM2 did for Vega). Add that to using 8nm instead of 7nm and there's your extra 100W.

Yeah, it's a power-hog, but if this turns out to be true it would still be the fastest consumer card by a fair margin (even with a 110-120% performance uplift, AMD would still lose).

 

[Krteq]

This one is huuuge

WCCFTech - NVIDIA GeForce RTX 3090 & RTX 3080 Ampere GA102 GPU Allegedly Pictured – Massive Die For Enthusiast Gaming Graphics Cards

 

[kurosaki]

That is going to cost at least one kidney, an arm and a leg to get hold of...

 

[n0x1ous]

This one is huuuge

WCCFTech - NVIDIA GeForce RTX 3090 & RTX 3080 Ampere GA102 GPU Allegedly Pictured – Massive Die For Enthusiast Gaming Graphics Cards

I need, I need. TW? Taiwan? Not Samsung?

 

[xpea]

Has Turing been faster ALU for ALU than RDNA1?

No. If so, then I do not see why next gen Nvidia and AMD GPUs would show different picture.

ALU to ALU is only a small part of the story. You always forget to mention that Turing is miles ahead of RDNA in terms of uarch. Turing is DX12 Ultimate generation with mesh shaders, VRS, sampler feedback, Ray Tracing and so on. Still, RDNA 7nm barely keeps up with 16nm Turing in power efficiency. RDNA2 will only catch up with Turing uarch, while Ampere will be on another level again, crushing RDNA2 in the important high-end features like Ray Tracing. And I don't even talk about DLSS 3, the elephant in the room...
All this talk about rasterization performance is to hide how far away AMD is in features. Hopefully, in few weeks, all be clear and AMD brigade will move goalpost and start to hype how RDNA3 MCM will be so much better

 

[Glo.]

ALU to ALU is only a small part of the story. You always forget to mention that Turing is miles ahead of RDNA in terms of uarch. Turing is DX12 Ultimate generation with mesh shaders, VRS, sampler feedback, Ray Tracing and so on. Still, RDNA 7nm barely keeps up with 16nm Turing in power efficiency. RDNA2 will only catch up with Turing uarch, while Ampere will be on another level again, crushing RDNA2 in the important high-end features like Ray Tracing. And I don't even talk about DLSS 3, the elephant in the room...
All this talk about rasterization performance is to hide how far away AMD is in features. Hopefully, in few weeks, all be clear and AMD brigade will move goalpost and start to hype how RDNA3 MCM will be so much better

Moving the goalposts, eh?

I love how you at the end talked about AMD brigade moving the goalpost, while in whole your post you moved the goalpost...

 

[AtenRa]

Are you honestly expecting Big Navi to have ~225% the performance of a 5700XT (which would be required for a clear win, if Nvidia can pull off a super dissapointing 40% uplift) ... and at @275W ?

From the latest Techpowerup review, the RTX2080ti is 40%-50% faster vs the RX5700XT at 1440p and 4K.
If RTX3090 is 50% faster vs RTX2080Ti then AMD needs to have a 125% faster card vs the RX5700XT to get even, not 225%

 

[Krteq]

I need, I need. TW? Taiwan? Not Samsung?

GP107 also manufactured at Samsung (14nm)

Spoiler

 

[Glo.]

Also, why do people expect Ampere to have 60% more performance than RTX 2080 Ti, in pure rasterization?

82 SMs is 20% more than RTX 2080 Ti had. Then you get higher memory bandwidth. And, at best 10% higher clock speeds(RTX 2080 Ti was able to clock to 1.95 Ghz).

Total performance of RTX 3090 should be around 40-50% higher than RTX 2080 Ti.

Where do people get that another 20-10% more? There is not that much that Nvidia can eek out from current GPU designs, WITHOUT vast redesign of the architecture. And that only will come with Chiplet based GPUs.

I can believe in that screen that was on previous pages only if it was With DLSS. But not pure Horsepower of the ALUs.

 

[Krteq]

And I don't even talk about DLSS 3, the elephant in the room...

You mean that MLID nonsense?