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.

 

[IntelUser2000]

Remember guys, GA100 chip has 108 SM's out of whole design having 128 SM's. How much will it draw power? 500W? And remember, this thing has HBM2! Where each stack draws 4W's of power. So we are talking about 20-30W for memory subsystem.

Probably more, if they really are moving to a 384-bit interface on Ampere gaming. And that won't be helped by the new process.

It's more than 30W for Ampere:

https://www.extremetech.com/wp-content/uploads/2016/02/NV-HB.png

60W for 1.25TB/s HBM2. At 1.6TB/s for Ampere we're talking 80W. 2TB/s HBM2, 125W.

 

[DooKey]

Why are we talking power draw? If the GA102 cards on TSMC process are the fastest cards available for gaming does it really matter? Are some saying that only HPC is on TSMC?

 

[Ajay]

Why are we talking power draw? If the GA102 cards on TSMC process are the fastest cards available for gaming does it really matter? Are some saying that only HPC is on TSMC?

Well, the current belief is that GA102 will be manufactured by Samsung, if it winds up being TSMC, I guess all bets are off. As to why it matters, people have bashed AMD GFX cards for being power hungry when explaining why NV is better (peft/watt). If NV's cards run at high power and temps, then they will be trashed for having poor perf/watt. In the end, what really matters is perf/$, so long as AMD's drivers don't suck again.

 

[Glo.]

Well, the current belief is that GA102 will be manufactured by Samsung, if it winds up being TSMC, I guess all bets are off. As to why it matters, people have bashed AMD GFX cards for being power hungry when explaining why NV is better (peft/watt). If NV's cards run at high power and temps, then they will be trashed for having poor perf/watt. In the end, what really matters is perf/$, so long as AMD's drivers don't suck again.

Its much more important than that.

Because if AMD GPUs will end up more efficient, and faster in rastrization AMD will achieve something unthinkable before.

They will make God, Nvidia, bleed.

And that is biggest win from mindshare perspective you imagine right now.

For which Nvidia can't allow. If their GPUs will be less efficient than AMD's they can make at least everything they can to win ultimate performance in certain segments.

This is why RTX 3080 is based on 102 die.

 

[Stuka87]

Thing is, even if nVidia is less efficient, many won't care. nVidia would have to be grossly worse for it to impact them.

 

[amenx]

People will care. Fermi left a bad taste in many peoples mouths. I will not be touching a 300w card. Even more so if its a 3080 102die which means it will be too expensive. Its why I'm skeptical re Glos speculation. If a 3080 102 die at best is $1000, what will the next card down be? People are not going to pay $650-750 for a 3070. I will skip this gen altogether. Therefore I hope the 102 die is called a 3080ti or 3090, because they have to leave some room below it for something that is above a 3070.

 

[Glo.]

People will care. Fermi left a bad taste in many peoples mouths. I will not be touching a 300w card. Even more so if its a 3080 102die which means it will be too expensive. Its why I'm skeptical re Glos speculation. If a 3080 102 die at best is $1000, what will the next card down be? People are not going to pay $650-750 for a 3070. I will skip this gen altogether. Therefore I hope the 102 die is called a 3080ti or 3090, because they have to leave some room below it for something that is above a 3070.

Its not my speculation. If you would read my posts, I was under the impression that RTX 3080 is going to be based on 104 or 103 die. But Nvidia appears to have cancelled it in favor of planting RTX 3080 on 102, per... Twitter user/Nvidia Leaker KopiteKimi:

https://twitter.com/x/status/1269239881512873984

 

[Glo.]

Thing is, even if nVidia is less efficient, many won't care. nVidia would have to be grossly worse for it to impact them.

Oh hell, people will care. The last last bastion of rhetorhic against AMD GPUs will fall. Bashing AMD for their GPUs being inefficient, and buying Nvidia cards, despite AMD GPUs being better value, will make people only look like complete hypocrites if they will not bash Nvidia, for the same thing, and buying Nvidia GPUs if the only thing Nvidia can do to compete with AMD is to lower prices below AMD offerings.

Its the same dilemma between RX 570 vs GTX 1050 Ti all over again but in reverse.

Will people buy more RX 6500 XT's, than direct competitor from Nvidia, if it is more efficient, than GTX 3050 Ti, despite being slower and costing the same?

 

[amenx]

Its not my speculation. If you would read my posts, I was under the impression that RTX 3080 is going to be based on 104 or 103 die. But Nvidia appears to have cancelled it in favor of planting RTX 3080 on 102, per... Twitter user/Nvidia Leaker KopiteKimi:

https://twitter.com/x/status/1269239881512873984

Fine, not your speculation, someone elses. The crux of my point left unanswered: what is below a 3080 102 that serves as the mid-range and not cost an arm and a leg?

 

[DXDiag]

You guys genuinely believe that those capabilities come as a free lunch from perspective of power draw? The very reason why Ampere GA100 chip is less efficient per watt compared to Volta in FP32 and FP64 are those increased AI capabilities in GA100 chip!

Consumer chips will cut down FP64 cores, decrease tensor cores count, each tensor core from Ampere is worth 4 of Volta or Turing, this will make them more agile and push their clocks.

considering that Samsungs process is LESS EFFICIENT than TSMC's 7 nm process

Nope, Gaming chips are 7nm not 8nm from Samsung.

The end result is 300W power draw on RTX 3080 based on 102 chip, and RTX 3090 drawing way past 300W of power. Why is it so hard to see this?

Maybe, performance is the thing that counts at the end of the day though.

Because the numbers, yes theoretical, show that Ampere is actually less efficient in FP32/64 per watt.

The TDP numbers are given for unlimited stable AI workloads, ie Tensor workloads since this is an AI optimized chip, the chip could consume just a 100w in FP32 workloads for all we care, you simply don't know that.

 

[DXDiag]

Remember guys, GA100 chip has 108 SM's out of whole design having 128 SM's. How much will it draw power? 500W? And remember, this thing has HBM2! Where each stack draws 4W's of power. So we are talking about 20-30W for memory subsystem.

Nope, each 16GB of HBM2 consumes about 50w, there is 40GB in A100, which means about 110W for HBM2.

The thing that I do not know was what clock speeds it was on. Also, considering the leak of 3080 Shroud I believe this GPU is closer to launch, on which I base my assumption that it is that 300W GPU talked here.

After you consistently said gaming chips are Q4?

Oh hell, people will care. The last last bastion of rhetorhic against AMD GPUs will fall. Bashing AMD for their GPUs being inefficient, and buying Nvidia cards, despite AMD GPUs being better value, will make people only look like complete hypocrites if they will not bash Nvidia, for the same thing, and buying Nvidia GPUs if the only thing Nvidia can do to compete with AMD is to lower prices below AMD offerings.

AMD is being bashed for a lot more than just power, there is bad drivers, lack of RT, lack of DX12 Ultimate features, lack of DLSS, lack of good decoder/encoder .. etc.

 

[IntelUser2000]

I think the PROBLEM for Ampere having worse perf/watt would be that they'll be criticized for not going with a pure rasterizer GPU and wasting resources on tensor/RT cores.

@DooKey If it requires 400W for 30% higher performance over a 300W Turing then you end up with zero perf/watt gain.

That's fine in the high end, for now. Eventually desktops will care. Maybe when next gen cards require 800W. But 400W is doable.

But notebooks and mainstream desktops will suffer if its based on that. Basically, anything except >$600 desktop cards.

 

[beginner99]

Why are we talking power draw? If the GA102 cards on TSMC process are the fastest cards available for gaming does it really matter?

It's important because when NV was more efficient than the competition, it was a very important metric.But always, once NV is less efficient, suddenly it shouldn't matter anymore.

The crux of my point left unanswered: what is below a 3080 102 that serves as the mid-range and not cost an arm and a leg?

The 3070 as always. The 3080 also won't sell for $1000 regardless what chip they put in it. It's what happens, wehn you have competition. NV can't price gouge that much anymore.

 

[AtenRa]

If those specs are correct NV marketing will go ballistic about RayTracing and DLSS 2.0/3.0 and completely skip raster performance.

 

[sontin]

Considering that CUT DOWN to 108 SM's chip draws 400W of power, why do you even believe Nvidia had time to optimize the physical design for N7 process, especially considering that is bog standard N7 process, and not like a version of 16 nm TSMC that Nvidia and TSMC called 12 nm FFN, for marketing purposes?

TSMC'S 7nm is in production since two years. Tape out of A100 was summer last year. So you claim that nVidia cant optimize the process prio the tape out? Sure.

 

[uzzi38]

NVIDIA GeForce RTX 3080 cooler could cost up to 150 USD alone - VideoCardz.com

NVIDIA has allegedly begun an investigation into the GeForce RTX 3080 leak. NVIDIA PG132 board to launch with three graphics cards Igor Wallossek from Igor’sLAB has new information on Saturday’s leak of RTX 3080 cooler design. The company has allegedly started an investigation. Two...

videocardz.com

Igor's Lab backing up Kimi on GDDR6X and 3080 and 3090Ti/SUPER (though adding in 3080Ti/SUPER) all using GA102 and providing more detailed power draw figures of 320/350W for TBP (which Nvidia rates as TDP anyway). Also the cooler alone costs $150.

 

[Tabalan]

Oh hell, people will care. The last last bastion of rhetorhic against AMD GPUs will fall. Bashing AMD for their GPUs being inefficient, and buying Nvidia cards, despite AMD GPUs being better value, will make people only look like complete hypocrites if they will not bash Nvidia, for the same thing, and buying Nvidia GPUs if the only thing Nvidia can do to compete with AMD is to lower prices below AMD offerings.

Its the same dilemma between RX 570 vs GTX 1050 Ti all over again but in reverse.

Will people buy more RX 6500 XT's, than direct competitor from Nvidia, if it is more efficient, than GTX 3050 Ti, despite being slower and costing the same?

Vast majority of people are not into tech and for them Nvidia mindshare will be sufficient reason to go for Ampere. How much [% wise] of gamers are even bothering to do some more advanced research when buying PC? I'd bet it's <<10%

 

[itsmydamnation]

Vast majority of people are not into tech and for them Nvidia mindshare will be sufficient reason to go for Ampere. How much [% wise] of gamers are even bothering to do some more advanced research when buying PC? I'd bet it's <<10%

explain Zen's DIY cpu share then?

 

[Krteq]

If those specs are correct NV marketing will go ballistic about RayTracing and DLSS 2.0/3.0 and completely skip raster performance.

I'm afraid so

 

[uzzi38]

If those specs are correct NV marketing will go ballistic about RayTracing and DLSS 2.0/3.0 and completely skip raster performance.

They won't skip it, it's still a notable bump gen on gen. But RTRT perf is definitely the star of the show for Ampere.

 

[Tabalan]

explain Zen's DIY cpu share then?

You know DIY is small market with highest % of tech enthusiasts (knowledge wise, not $$$ wise)? More important market is for example laptops. When Intel had CPU shortage DIY was least import for them.

 

[Veradun]

If those specs are correct NV marketing will go ballistic about RayTracing and DLSS 2.0/3.0 and completely skip raster performance.

Will they suggest reviews with just three games?

 

[Tabalan]

Will they suggest reviews with just three games?

According to rumours DLSS 3.0 is supposed to work with every game with TAA implemented. They'll just suggest using those games and Ampere GPUs will run circles around Turing/Pascal.

 

[FiendishMind]

Wasn't Kopite adamant GA102 taped out in March? Is it even possible for Nvidia to have gotten silicon back so quickly?

 

[Krteq]

According to rumours DLSS 3.0 is supposed to work with every game with TAA implemented. They'll just suggest using those games and Ampere GPUs will run circles around Turing/Pascal.

What rumors? Any link?