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.

 

[sontin]

I dont give him clicks. The 2080TI uses up to 22% more power for ~40% more performance over the 1080TI.
A100 PCIe uses 250W for 2,25x more performance over V100 PCIe.

 

[jpiniero]

Sure, 400W. So for 40% more performance over the 2080TI Ampere needs 45% more power.

Yeah, right.

Everything looks about where you expect it. nVidia is hitting their targets. The power draw is just because it's on SS8 instead of SS7.

Now by being on 7+ AMD could theoretically do a Yuge Navi (ie: like 600+ mm2) and there would be nothing nVidia could do about it. That doesn't mean they will, but it's a possibility.

 

[Glo.]

Meh. Honestly sounds like MLID just put together leaks like kopite’s and added a few things. I can’t take him seriously given his record.

Or simply kopite and MLID are reporting about real engineering samples being tested by Nvidia engineers/Nvidia partners Engineers.

 

[GodisanAtheist]

You guys don't understand, the 3080 NEEDS to be good and cheap... so that I can pick up a 2080S or 2080Ti on firesale on the second hand market.

As such, I believe the $500 3080 will be 1 MILLION % better than the 2080TI, which I can take off your hands for $400 before the market on used 2080's crashes thanks to the impending new gen release.

 

[Stuka87]

Sure, 400W. So for 40% more performance over the 2080TI Ampere needs 45% more power.

Yeah, right.

400W is the overclocked value, as clearly stated. When over clocking, power usage is NEVER linear with performance increases. This is very much common knowledge. 40% more performance for only 45% more power would actually be extremely good.

I dont give him clicks. The 2080TI uses up to 22% more power for ~40% more performance over the 1080TI.
A100 PCIe uses 250W for 2,25x more performance over V100 PCIe.

That 2.25x is in a very specific work load that has zero impact on gaming. According to nVidia FP32 goes from 16.4TF to 19.5 for Volta to Ampere PCIe.

 

[MrTeal]

I dont give him clicks. The 2080TI uses up to 22% more power for ~40% more performance over the 1080TI.
A100 PCIe uses 250W for 2,25x more performance over V100 PCIe.

Turing on 12FFn got an 18% increase in PPW relative to Pascal on 16nm. I don't know if we'll see 40% going from Turing to Ampere, but it doesn't seem out of line to me. Even the 1080 had 60% increase in PPW relative to the 980 at launch going from 28nm to 16nm, and that was a more substantial process jump than this.

A100 might have have 2.25x performance over V100 in specific tasks, but it certainly doesn't in general floating point performance. Until Ampere is released as a graphics product we won't be able to really compare how they perform in traditional raster games, but PCIe A100 gets 38% more SP FLOPS than V100 in the same power budget.

Now in RT (or with a new version of DLSS) applied I'd expect that increase to be much larger. If Ampere makes turning ray tracing on possible without a massive performance hit, it's going to make any kind of straight FPS comparison murky since you're not really comparing like for like.

 

[sontin]

8nm is a full node shrink from TSMC's 16nm process. And A100 delivers twice the performance in "general floating point" applications (FP16 and FP64). A100 is twice as efficient as Turing T4 in inference workloads, too.

 

[MrTeal]

8nm is a full node shrink from TSMC's 16nm process. And A100 delivers twice the performance in "general floating point" applications (FP16 and FP64). A100 is twice as efficient as Turing T4 in inference workloads, too.

Source? Nvidia lists 9.7 TF for FP64, compared with 7TF for V100.

 

[alcoholbob]

8nm is a full node shrink from TSMC's 16nm process. And A100 delivers twice the performance in "general floating point" applications (FP16 and FP64). A100 is twice as efficient as Turing T4 in inference workloads, too.

A100 is on 7nm TSMC, not 8nm Samsung. Also Samsung's chips have always had way more power consumption than TSMCs, whether you compare cell phone chips or GPUs when the same chip is produced by both foundries. Plus 40% more performance for 40% more power consumption is a significant improvement. You realize squeezing out an extra 15% of performance from an AIB 2080 Ti overclocking takes like 60-70% more power consumption?

 

[sontin]

Samsung's 8nm process is more efficient than TSMC 16nmFF (even their 10nmFF process): https://www.anandtech.com/show/12520/the-galaxy-s9-review/6

We went over it a month ago.

 

[Tup3x]

I am sure Intel was thinking that exact same thing two years ago. Ryzen 1 was a big jump over prevision FX chips, but didn't beat Intel. Ryzen 2 came out, and well, we all know how that is going. Navi 1 was step one (akin to Ryzen 1). Navi 2 is the next big jump. Obviously its quite possible that nVidia will also have a big jump and still come out ahead. But to say AMD will only get close in a best case scenario is shortsighted.

Definitely not the same thing. NVIDIA is not going to be stuck with same process. Judging by all info AMD has said about RDNA2 there really is no reason to believe that it's suddenly going to wipe the floor with Ampere. 50% performance/w improvement isn't enough to expect AMD to suddenly be ahead. They should finally be competitive in €€€€ tier cards or at least offer more than just mid and low tier cards. I'd except NVIDIA to have fastest card this time too. I hope that RDNA2 would make it much harder for NVIDIA and force them to lower price (I fear that AMD just ends up rising theirs...).

I don't expect either to be absolutely in another class this time at least in rasterisation performance. Ray tracing is going to be another thing entirely... I expect Ampere to be quite a bit ahead (after all, NVIDIA had quite some time to make improvements).

 

[Konan]

I don't expect either to be absolutely in another class this time at least in rasterisation performance. Ray tracing is going to be another thing entirely... I expect Ampere to be quite a bit ahead (after all, NVIDIA had quite some time to make improvements).

This.

Sounds like things will be competitive and I welcome that. I also believe that there will not be that much significance in TDP between them. e.g. 20-30W is nothing.
Lastly, software is king... DLSS, drivers etc. can be game changers and I see one company firmly above the other here, although I genuinely hope that this area becomes more competitive too!

 

[FaaR]

I can’t take him seriously given his record.

I think this guy is full of crap, honestly. He spouts loads of nonsense apparently all the time from what I can tell; it's gotta be pretty bad considering I've only ever watched two of his videos (this one being one of them, and I didn't watch more than the first minute or so.)

I wouldn't bet anything of any value on anything this guy says.

 

[AtenRa]

8nm is a full node shrink from TSMC's 16nm process. And A100 delivers twice the performance in "general floating point" applications (FP16 and FP64). A100 is twice as efficient as Turing T4 in inference workloads, too.

And this has to do what with Gaming performance and gaming performance efficiency ???

 

[Glo.]

50% performance/w improvement isn't enough to expect AMD to suddenly be ahead.

50% performance/watt uplift means two things.

RX 5700 XT performance at 115W of power consumed, or RTX 2080 Ti performance at 230W of power consumed.

If AMD was able to bring 80 GPU to around 2 GHz mark at 290W of power, that GPU will be 50-60% faster than RTX 2080 Ti, based on doubling the CU count, higher IPC, higher clock speeds, higher memory bandwidth. And that 80 CU GPU? It may be only cut down part, from full die of 96 CUs. May be. To be confirmed.

Current rumors say, that RTX 3080 Ti is 40% faster than RTX 2080 Ti, and Nvidia is desperate to push it higher. Current rumors say, that Nvidia cancelled GA103 chip, and put RTX 3080 on GA102 die.

Why would they do this IF THEIR architecture would be competitive, against RDNA2? It isn't, which is why Nvidia had to retool the designs, for higher clock speeds, and threw out the window the efficiency. Hence why they rumors about 300-350W of power drawn.

Turing's IPC already was exactly the same as RDNA1. What if RDNA2 increased it by 15%, and next gen gaming cards from Nvidia achieve below 10% IPC increase? You get lower clocking GPU architecture, that has lower IPC, with worse efficiency, than direct competitor. Its not a problem for Nvidia to lose mindshare. They can afford it.

The problem for them with this is that AMD Radeon group IS GAINING MINDSHARE, with this. And in this business, its far more important. Its easy to be "great" when your competitor cannot deliver. Its way harder to compete with your competitor being able to compete and having something you do not have. CPUs, with which they can sell their products. Look at freaking Intel, what has happened with them. 7 nm EUV fiasco for Nvidia is costing them more, and will cost them far more in the long that anyone here anticipates.

Why? Because AMD has made god, Nvidia, bleed. AMD, or rather Nvidia's own stupidity, has made them "mortal".

The good news from this is that few months ago one of my sources told me this: "if you believe Nvidia will be wiped out, be unable to compete - you are ******* stupid."

Compete - yes. But they will have inferior products.

I personally do not care who wins next round. Based on what I know, nothing that would interest me in the GPU space will come before 2021, and I don't want to wait to that day for new GPU.

 

[BenSkywalker]

So all these rumors you all have been quoting says the 3080ti is a 375 watt part. We see benchmarks showing an upcoming GPU from nVidia roughly 30% faster than the 2080ti and most of the people in this thread say that had to be the 3080ti, not the 3080.

Now the same rumor chain says that is a 300 watt part and everyone instantly drops their 375 watt rumor because it doesn't fit with their previous narrative.

You all are doing a lot of mental gymnastics to create a rumor that gives AMD the best chance- you ignore your own rumors in the blink of an eye when it looks dangerous to your perception of where AMD is going to land.

This new routine does imply the Koptite was wrong *again* saying it was a 375 watt part- you all have made this kid sound like the least reliable rumor creator ever.

 

[MrTeal]

The problem for them with this is that AMD Radeon group IS GAINING MINDSHARE, with this. And in this business, its far more important. Its easy to be "great" when your competitor cannot deliver. Its way harder to compete with your competitor being able to compete and having something you do not have. CPUs, with which they can sell their products. Look at freaking Intel, what has happened with them. 7 nm EUV fiasco for Nvidia is costing them more, and will cost them far more in the long that anyone here anticipates.

Why? Because AMD has made god, Nvidia, bleed. AMD, or rather Nvidia's own stupidity, has made them "mortal".

The good news from this is that few months ago one of my sources told me this: "if you believe Nvidia will be wiped out, be unable to compete - you are ******* stupid."

Compete - yes. But they will have inferior products.

That's a pretty big stretch. Other than some internet speculation, AMD hasn't made god bleed in years. The rumour mill is working overtime with hypotheticals of the upcoming gen, but AMD hasn't been really competitive at the high end since Hawaii. Even then, AMD barely made a dent in Nvidia's market share. I'd love to see big Navi be really competitive for the absolute performance crown, but even if it is AMD will need multiple generations of superior products before they really start to bleed customers away from Nvidia in a way that harms them long term.

 

[DiogoDX]

Nvidia will not lose the performance crown. They already show with the GTX480 that they are willing to trow the power to beat AMD by 10-15%.

 

[Glo.]

That's a pretty big stretch. Other than some internet speculation, AMD hasn't made god bleed in years. The rumour mill is working overtime with hypotheticals of the upcoming gen, but AMD hasn't been really competitive at the high end since Hawaii. Even then, AMD barely made a dent in Nvidia's market share. I'd love to see big Navi be really competitive for the absolute performance crown, but even if it is AMD will need multiple generations of superior products before they really start to bleed customers away from Nvidia in a way that harms them long term.

Its not a matter of making Nvidia bleed on high end. Current picture of the information we got on hand says that AMD will have better chips, from top, to bottom, with new generation. The only place where they might not have better products is... low-end. 107 and 106 might be better than... Navi 10 based GPUs.

 

[BenSkywalker]

Current picture of the information we got on hand says that AMD will have better chips, from top, to bottom, with new generation.

Your sources have proven to be hilariously wrong on every single point, you keep choosing to ignore how far off what they say now is to what they were saying mere months ago and it goes *way* beyond moving the goalposts, you've changed the freaking sport multiple times.

No consumer Ampere- wrong
All parts SS7nm - Now he says this is wrong
No consumer parts this year- now he says this is wrong
Forced retape of everything to SS8nm magically makes the parts come 2 quarters earlier...
375 watt power usage
300 watt benches leak- that 375 watt rumor was wrong

With all of this, you all keep clinging to the SS8nm- he's been wrong on every other point according to what you all are saying now, and yet you are convinced this one caveat must be true.

If we go back over the last several generations, the more effort on the forums to hype new parts, the worse the part ends up being, you are collectively making this one sound very, very bad.

 

[uzzi38]

There's also the notably shorter cadence (12-18 months compared to 24-27 months) that AMD is following now.

Even if AMD don't take the crown this generation and fall ever so slightly behind, they have a nearly guaranteed 6 months ahead of the competition. And if, say hypothetically, RDNA3 was a faster follow up than RDNA2 was and you're looking at closer to the 12 month side of things, well then you've got about a full year with AMD set to bring about RDNA4 within half a year before Nvidia releases their 5nm counter to RDNA3.

What people here don't seem to realise is that it's not RDNA2 that's the real danger to Nvidia. The next generation will be close but that's all. It's AMD's swift cadence and +50% perf/W each generation promise that is a real danger.

A cadence shift like this is not easy at all. Rumours that Nvidia are putting orders out for 5nm capacity in Q4 2021 tell you that they'll begin volume production then for what is most likely to be the datacentre card. Consumer cards will likely follow after.

Based off that I'd estimate that Nvidia would be capable of reigning in their cadence by ~3 months by this at best. Imagine that situation. The best case scenario for Nvidia is to be significantly behind by 3 months. The worst case? Closer to a year.

RDNA2 is just the beginning - honestly I don't entirely care if it pulls ahead of Ampere or not because it's gonna be a photo finish either way. The main reason RDNA2 is interesting is because it's a clear statement of whether or not their targets mean anything. That's all.

Also, as a side note, can someone fill me on on this 300W thing? I'm completely out of the loop I guess.

 

[Glo.]

Your sources have proven to be hilariously wrong on every single point, you keep choosing to ignore how far off what they say now is to what they were saying mere months ago and it goes *way* beyond moving the goalposts, you've changed the freaking sport multiple times.

No consumer Ampere- wrong
All parts SS7nm - Now he says this is wrong
No consumer parts this year- now he says this is wrong
Forced retape of everything to SS8nm magically makes the parts come 2 quarters earlier...
375 watt power usage
300 watt benches leak- that 375 watt rumor was wrong

With all of this, you all keep clinging to the SS8nm- he's been wrong on every other point according to what you all are saying now, and yet you are convinced this one caveat must be true.

If we go back over the last several generations, the more effort on the forums to hype new parts, the worse the part ends up being, you are collectively making this one sound very, very bad.

I said this once to Sontin. I will repeat it for you.

You should work on your reading comprehension skills, mate.

 

[AtenRa]

The only reason AMD have a chance against Ampere is because they havent heavily invested in RayTracing and Tensors like NVIDIA.
The notion that AMD has a clear edge because of the 8nm vs 7nm is not valid.

 

[BenSkywalker]

It's AMD's swift cadence and +50% perf/W each generation promise that is a real danger

Let's see them do it once without a full node drop.

"Poor Volta"

Three years waiting on that claim, but now they are trustworthy?

 

[Stuka87]

The only reason AMD have a chance against Ampere is because they havent heavily invested in RayTracing and Tensors like NVIDIA.
The notion that AMD has a clear edge because of the 8nm vs 7nm is not valid.

RDNA 2 has RTRT, but AMD went about it in a very different way. Each shader had the functionality built into it. Rather than having a big chunk of the GPU dedicated to hardware that can only do RTRT. What this means for performance, we don't know. But in theory, it means the AMD GPU's could be more balanced. Which was one of Turing's failings. Not nearly enough RT hardware to let the card run at regular rasterized speeds.

EDIT: Its part of the TMU, not the shader. My mistake.