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.

 

[uzzi38]

No, Navi 10 uses N7FF

Sent from my SM-N975F using Tapatalk

It's N7P as per AMD themselves:

Radeon RX 5700: Navi and the RDNA Architecture

A look at AMD's Radeon RX 5700 GPU built on a 7-nanometer process based on the new Navi microarchitecture and RDNA graphics architecture.

fuse.wikichip.org

 

[Glo.]

I'd love for Navi 2 to be competitive or even better, but I'm used to being disappointed by AMD gpu's after owning Fury X, Vega 64, and Radeon VII

I will put it this way, once again.

Ask yourself, why did Nvidia clocked their GPUs to hell, sacrificing efficiency, and why they have put RTX 3080 SKU on 102 die, instead of, traditionally, 104.

Because otherwise their products would staright up lose against AMD.

 

[uzzi38]

You know what I found especially hilarious as per a certain someone's post yesterday?

He mentionned that this generation feels like Fury vs Maxwell all over again. And I honestly have to agree. I just think it's more likely the other way around this time.

We have one GPU lineup that has clocked pushed to the absolute wall, with expensive, brand new memory that will seriously bump up prices on it's own.

Then we have another GPU lineup that's pulled off a significant perf/W uplift despite being on the same node primarily due to architectural improvements. As a result it performs better at the same power, and it also uses a much more sane memory config (probably on the last. We don't know for certain anyway).

I'll let you guys fill in the blanks on what's what

 

[Glo.]

Power of engineering versus power of arrogance.

That is how this gen can be summed up.

 

[Kenmitch]

Power of engineering versus power of arrogance.

That is how this gen can be summed up.

If rumors hold true Nvidia will have the BIGGER card. BIGGER is better!

 

[sontin]

You know what I found especially hilarious as per a certain someone's post yesterday?

He mentionned that this generation feels like Fury vs Maxwell all over again. And I honestly have to agree. I just think it's more likely the other way around this time.

We have one GPU lineup that has clocked pushed to the absolute wall, with expensive, brand new memory that will seriously bump up prices on it's own.

Then we have another GPU lineup that's pulled off a significant perf/W uplift despite being on the same node primarily due to architectural improvements. As a result it performs better at the same power, and it also uses a much more sane memory config (probably on the last. We don't know for certain anyway).

I'll let you guys fill in the blanks on what's what

And all that without knowing Ampere's gaming performance. Not bad. BTW: Why is Sony producing the biggest console ever? 10 TFLOPs soundnt be too hard to cool, huh?

 

[DDH]

It's N7P as per AMD themselves:

Radeon RX 5700: Navi and the RDNA Architecture

A look at AMD's Radeon RX 5700 GPU built on a 7-nanometer process based on the new Navi microarchitecture and RDNA graphics architecture.

fuse.wikichip.org

Ok thanks, although I don't see any citation linking to where AMD provided information

Sent from my SM-N975F using Tapatalk

 

[Kenmitch]

And all that without knowing Ampere's gaming performance. Not bad. BTW: Why is Sony producing the biggest console ever? 10 TFLOPs soundnt be too hard to cool, huh?

The logical thing to do is wait for both offerings to be reviewed.

Would you go to a gun fight and leave your ammo behind?

 

[uzzi38]

Ok thanks, although I don't see any citation linking to where AMD provided information

Sent from my SM-N975F using Tapatalk

It's on one of the provided slides.

PS: You can remove the auto-generated watermark left by Tapatalk in the end of each post.

 

[uzzi38]

And all that without knowing Ampere's gaming performance. Not bad. BTW: Why is Sony producing the biggest console ever? 10 TFLOPs soundnt be too hard to cool, huh?

Because their last was a complete jet engine and everybody hated it. Now it's time for heatsink galore.

Ever heard what a PS4 sounds like in-game? I certainly could from halfway down the road.

 

[MrTeal]

The logical thing to do is wait for both offerings to be reviewed.

Would you go to a gun fight and leave your ammo behind?

Yeah, there are a lot of very authoritative statements in this thread on how exactly this generation will play out when we have nothing but speculation and leaks on what performance, power and pricing will be from either camp.

 

[jpiniero]

Ask yourself, why did Nvidia clocked their GPUs to hell, sacrificing efficiency, and why they have put RTX 3080 SKU on 102 die, instead of, traditionally, 104.

Originally it was going to be GA103 (3840 cores). That they bumped it to GA102 may be more due to not wanting to port the GA103 die back from SS7. That they bumped it to 4352 is probably due to AMD.

 

[Stuka87]

And all that without knowing Ampere's gaming performance. Not bad. BTW: Why is Sony producing the biggest console ever? 10 TFLOPs soundnt be too hard to cool, huh?

Nobody said it was hard to cool. We already know what the power consumption is for both consoles, neither uses much power for an entire system with an 8 core Zen2 with GPU.

The size has more to do with all the complaints about PS4's (especially Pro's) being loud. Making the console lightly larger allows for quieter cooling.

 

[moonbogg]

Yeah, I'm not paying $800 for a downgrade in Vram amount from a 1080Ti. I'm also skeptical about Ampere's ability to simply not break in a year or 2 due to extreme current draw through the card. I don't like what I'm seeing so far.

 

[maddie]

I'm quite a bit mystified by the enormous apparent power draw of the 3090 & 3080 models and the leaked performance improvement, all done on a supposedly very superior 8nm SS node relative to TSMC 12nm.

The consensus belief is that 12nm is a slightly improved 16nm.

What if this is not quite true?

Three reasons give some doubt to the accepted wisdom.
1) There was a rumor early on that Nvidia had a customized 12nm process developed for their use case.
2) GloFlo has developed a 12nm node (12LP+) which has 40% less power consumption than their 14LPP.
3) TSMC is several years ahead of GloFlo.

What if Pascal and Turing were the beneficiaries of a much better node than is commonly believed and the progression to SS 8nm is a lot less of a efficiency jump than we think?

It clears up in my mind the issue I started with. The (power draw/performance/node efficiency) triangle.

Everyone keeps saying 12nm is really a marketing 14nm. What if that belief is all wrong as GloFlo shows is possible?

 

[uzzi38]

Yeah, I'm not paying $800 for a downgrade in Vram amount from a 1080Ti. I'm also skeptical about Ampere's ability to simply not break in a year or 2 due to extreme current draw through the card. I don't like what I'm seeing so far.

Unfortunately I don't think you'll be seeing less than $900 minimum for more than 11GB VRAM regardless of vendor.

Honestly better off not basing any decisions on that one front alone.

 

[sontin]

I'm quite a bit mystified by the enormous apparent power draw of the 3090 & 3080 models and the leaked performance improvement, all done on a supposedly very superior 8nm SS node relative to TSMC 12nm.

The consensus belief is that 12nm is a slightly improved 16nm.

What if this is not quite true?

TSMC 12nm is just an optimized 16nm process. You can use Pascal GP107 on Samsungs 14nmLPP as a reference. Ampere wouldnt be really better from a efficiency point.

 

[Tup3x]

I hate this silly season... At least all should become clear soon when it comes to NVIDIA. Then the massive hype for RDNA2 starts (like usual).

 

[Glo.]

I'm quite a bit mystified by the enormous apparent power draw of the 3090 & 3080 models and the leaked performance improvement, all done on a supposedly very superior 8nm SS node relative to TSMC 12nm.

The consensus belief is that 12nm is a slightly improved 16nm.

What if this is not quite true?

Three reasons give some doubt to the accepted wisdom.
1) There was a rumor early on that Nvidia had a customized 12nm process developed for their use case.
2) GloFlo has developed a 12nm node (12LP+) which has 40% less power consumption than their 14LPP.
3) TSMC is several years ahead of GloFlo.

What if Pascal and Turing were the beneficiaries of a much better node than is commonly believed and the progression to SS 8nm is a lot less of a efficiency jump than we think?

It clears up in my mind the issue I started with. The (power draw/performance/node efficiency) triangle.

Everyone keeps saying 12nm is really a marketing 14nm. What if that belief is all wrong as GloFlo shows is possible?

That 12 nm FFN have had allowed for specific physical design optimizations, that would work only for Nvidia designs.

8 nm Process does not have them, because Nvidia did not had the time to develop those physical design optimizations.

This is the reason why there is not that huge jump from 12 nm FFN to 8 nm LPP from Samsung.

If Nvidia would have the time to develop those physical design optimizations expect that that 2.1 GHz clock speed on RTX 3090 would be hit with 275W of power, instead of 375W's.

 

[sontin]

3090 doesnt clock up to 2,1Ghz. It is 1700Mhz from the leakers. And A100 is 1430Mhz and 100MHz less than V100. So i guess nVidia didnt have enough time, too?

 

[MrTeal]

Yeah, I'm not paying $800 for a downgrade in Vram amount from a 1080Ti. I'm also skeptical about Ampere's ability to simply not break in a year or 2 due to extreme current draw through the card. I don't like what I'm seeing so far.

It's a weird place. The obvious upgrade would be the 3090, but 24GB is excessive for gaming unless you're using mods just for the sake of trying to max things out and the reported $1400 is a massive ask. I don't know what the cost adder is for the extra 12GB of GDDR6X, but I'd rather have the 3090-class die and 12GB if it knocks $200 off the MSRP.

 

[raghu78]

Navi is on 7nm and failed against Pascal. Maybe we should think about that. And yet we have people here who are claiming that AMD will make Navi 2 more than twice as efficient as Navi 1. How should this work?

You missed the Xbox Series X power numbers as claimed by Microsoft. 52 CUs at 1825 Mhz for 12.15 TF at same power consumption as Xbox One GPU. The GPU in itself is <=100w. With 16GB GDDR6 memory its still roughly 130-135w. RDNA2 brings the best physical design methodologies used by the Zen CPU core team. The numbers are there for everyone except the most stubborn Nvidia supporters to see.

Look at the rumoured power numbers for GA104 with 2 SKUs - RTX 3070 Ti (3072 CC) , RTX 3070 (2944 CC)

https://twitter.com/x/status/1280311636629192705

https://twitter.com/x/status/1280424859093184512

250w for roughly 12.3 TF based on a 2 Ghz boost clock.

The writing is on the wall. Samsung 8nm is no match for TSMC N7P. Nvidia underestimated AMD and bet on the wrong foundry and process. Ampere is going to be Nvidia's worst perf/watt product generation after Fermi. The 21 day hype on Geforce twitter account and videos about their cooler design is just damage control.

 

[Glo.]

3090 doesnt clock up to 2,1Ghz. It is 1700Mhz from the leakers. And A100 is 1430Mhz and 100MHz less than V100. So i guess nVidia didnt have enough time, too?

Its official clock. And we know that RTX 2080 Ti official clocks were lower than during gaming .

No need to play the damage limitation game. Nobody here is stupid, and cannot look for reviews and get something like this:

NVIDIA GeForce RTX 2080 Ti Founders Edition 11 GB Review

NVIDIA debuted its Turing graphics architecture today, straightaway with the flagship RTX 2080 Ti. This card packs the promise of real-time ray tracing at 4K UHD, besides huge gains in performance. NVIDIA also put out its best cooler design since TITAN, commanding a very high price for some very...

www.techpowerup.com

3D Load	1607-1950 MHz​	1750 MHz​	0.843-1.068 V​

That 2.1 GHz clock that we talk is going to be the same, new 1950 MHz boost, during the game.

 

[Glo.]

It's a weird place. The obvious upgrade would be the 3090, but 24GB is excessive for gaming unless you're using mods just for the sake of trying to max things out and the reported $1400 is a massive ask. I don't know what the cost adder is for the extra 12GB of GDDR6X, but I'd rather have the 3090-class die and 12GB if it knocks $200 off the MSRP.

Its funny.

Everybody talks about Nvidia price/performance/efficiency = value offering, but nobody is willing to even discuss, the obvious direction - and picking the competition this time around if they provide better value.

Maybe after all people really wanted AMD to be competitive, only to be able to buy Nvidia GPUs cheaper?

 

[JujuFish]

Navi is on 7nm and failed against Pascal. Maybe we should think about that. And yet we have people here who are claiming that AMD will make Navi 2 more than twice as efficient as Navi 1. How should this work?

Why does the idea of AMD pulling out a winner upset you so much? What's the downside? It's not going to magically make GeForce cards run worse, and if anything it would allow for a potential price war, benefiting us, the consumers.