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.

 

[Mopetar]

If these 2 aren't for a separate compute uarch, like Volta.

Nah, totally gaming cards. Everyone knows you want at least 47 GB of VRAM these days so you don’t get burned and hit the wall when games start using more than 32 GB.

 

[Bouowmx]

6144-bit HBM2 (6 * 8 GB) confirmed. Or less likely, still 4096-bit (4 * 12 GB).

 

[lobz]

Nah, totally gaming cards. Everyone knows you want at least 47 GB of VRAM these days so you don’t get burned and hit the wall when games start using more than 32 GB.

I didn't mean the cards themselves right now But everyone's talking about Ampere right now, like they did about Volta, which turned out to be Pascal It also wasn't _that_ long ago.

 

[Det0x]

I didn't mean the cards themselves right now But everyone's talking about Ampere right now, like they did about Volta, which turned out to be Pascal It also wasn't _that_ long ago.

?
Maybe i misunderstand you, but Pascal got release before Volta

Pascal release date June 10, 2016
Volta release date December 7, 2017
Turing release date September 20, 2018

The Turing architecture is pretty much Volta + RT cores (Volta had tensor cores from the start)

We got the gaming version of Volta architecture 10 months later then V100,

 

[lobz]

?
Maybe i misunderstand you, but Pascal got release before Volta

Pascal release date June 10, 2016
Volta release date December 7, 2017
Turing release date September 20, 2018

The Turing architecture is pretty much Volta + RT cores (Volta had tensor cores from the start)

We got the gaming version of Volta architecture 10 months later then V100,

You totally misunderstood me, yes. Just until right before release day Videocardz leaked the Pascal name, for 1,5 years everybody on every forum was convinced it was going to be Volta, which turned out to be a computing-exclusive uArch for later instead. Also nobody knew Pascal would be Maxwell on steroids - pretty damn good steroids though, Pascal was excellent - but Maxwell was a true wonderchild in comparison, for a 28nm design.
When I see the same happening with the Ampere name, I just safely assume that nobody here has any idea what the next gen will bring, whether it turns out to be named Ampere or not.

NVIDIA keeps its cards successfully the very closest to themselves that I've ever seen in this industry.

 

[eek2121]

I'm standing behind my thoughts that:

1. Ampere is likely data center only.
2. 7nm/7nm EUV NVIDIA parts are still a ways off, likely Q4.
3. The performance bump won't be as big as many people think. Likely 30% or so increase (excluding ray tracing).
4. AMD's unreleased next gen GPUs are built to compete with NVIDIA's next gen.

 

[xpea]

I'm standing behind my thoughts that:

1. Ampere is likely data center only.
2. 7nm/7nm EUV NVIDIA parts are still a ways off, likely Q4.
3. The performance bump won't be as big as many people think. Likely 30% or so increase (excluding ray tracing).
4. AMD's unreleased next gen GPUs are built to compete with NVIDIA's next gen.

1. Yes and no (Nvidia datacenter and gaming/visualisation products are based on same architecture with only minor adjustments to match their target market)
2. Absolutely not. Big Ampere for datacenter has been tape out one year ago with drivers running since last October. The first 2 gaming dies have been tape out more than 6 months ago...
3. No. Regular shader performance increase will be higher than Turing vs Pascal
4. Yes it's AMD wish. But in reality not really...

Quote all of this and we talk again after GTC

 

[uzzi38]

1. Yes and no (Nvidia datacenter and gaming/visualisation products are based on same architecture with only minor adjustments to match their target market)
2. Absolutely not. Big Ampere for datacenter has been tape out one year ago with drivers running since last October. The first 2 gaming dies have been tape out more than 6 months ago...
3. No. Regular shader performance increase will be higher than Turing vs Pascal
4. Yes it's AMD wish. But in reality not really...

Quote all of this and we talk again after GTC

1. Yes.

2. Yes and no. Original tapeouts at Samsung did happen, but Nvidia are no longer moving forwards with Samsung for everything but low end.

3. Maybe. I don't know.

4. Better than you'd think given where they're at now. Far better than you'd think.

Still a fair bit behind, but yearly cadence will fix that.

 

[RetroZombie]

I'm surprised with the rumors of so many cores units (7552 cuda cores), was expecting they would be hitting some kind of wall, and they would go for less but bigger and more powerful units instead.

 

[DXDiag]

The performance bump won't be as big as many people think. Likely 30% or so increase (excluding ray tracing).

30% on a new node? that's not going to happen, even a 2080Ti is 40% faster than a 1080Ti on the same node! Expect at least 70% uplift from Ampere, that's the usual NVIDIA uplift on a new node.

Look at the leaked 7550 core chip, that's a very large ALU increase over Turing. Expect a Titan 3000 to offer substantial performance over Titan RTX.

 

[Thala]

30% on a new node? that's not going to happen, even a 2080Ti is 40% faster than a 1080Ti on the same node! Expect at least 70% uplift from Ampere, that's the usual NVIDIA uplift on a new node.

Look at the leaked 7550 core chip, that's a very large ALU increase over Turing. Expect a Titan 3000 to offer substantial performance over Titan RTX.

Nvidia wont make their cards faster than they have to. It is never a question of what is technically possible.
In particular if they decide to keep the current price brackets intact, you will not see 70% performance increase - not even close.

 

[Stuka87]

30% on a new node? that's not going to happen, even a 2080Ti is 40% faster than a 1080Ti on the same node! Expect at least 70% uplift from Ampere, that's the usual NVIDIA uplift on a new node.

Look at the leaked 7550 core chip, that's a very large ALU increase over Turing. Expect a Titan 3000 to offer substantial performance over Titan RTX.

1080Ti and 2080Ti are not on the same node. One is 16nm, one is 12nm.

 

[JasonLD]

1080Ti and 2080Ti are not on the same node. One is 16nm, one is 12nm.

TSMC 12nm is just another variant of their 16nm.

 

[beginner99]

30% on a new node? that's not going to happen, even a 2080Ti is 40% faster than a 1080Ti on the same node! Expect at least 70% uplift from Ampere, that's the usual NVIDIA uplift on a new node.

Dream on. Look at the die size of a 2080TI (GT102 chip). It's huge. 7nm and it's variants simply are too expensive for now to make such huge dies possible (which are needed for your suggested performance uplift especially if you consider even more added ray-tracing hardware.) Plus many have said that the gains from 7nm aren't that great really in terms of heat and power-use. 70% won't happen for shader gaming performance. Only area were it could happen is in the AI/deeplearning instructions or ray tracing performance.

 

[DXDiag]

Nvidia wont make their cards faster than they have to. It is never a question of what is technically possible.
In particular if they decide to keep the current price brackets intact

I disagree, they did this very same jump from GTX 580 to 780Ti, and then from 980Ti to 1080Ti.

Dream on. Look at the die size of a 2080TI (GT102 chip). It's huge.

Turing will be significantly smaller on 7nm, look at the size of the leaked Ampere chips, it's also huge.

Plus many have said that the gains from 7nm aren't that great really in terms of heat and power-use.

AMD seems to have gained significantly from 7nm in the GCN to RDNA transitions. I expect the same thing to happen to NVIDIA as well.

 

[Thala]

I disagree, they did this very same jump from GTX 580 to 780Ti, and then from 980Ti to 1080Ti.

Look, your fallacy is assuming everything else being equal, which is not the case. The cost per gate does not decrease remotely as much when going to 7nm compared to the nodes you are referring to with your examples.

 

[uzzi38]

I disagree, they did this very same jump from GTX 580 to 780Ti, and then from 980Ti to 1080Ti.

Turing will be significantly smaller on 7nm, look at the size of the leaked Ampere chips, it's also huge.

AMD seems to have gained significantly from 7nm in the GCN to RDNA transitions. I expect the same thing to happen to NVIDIA as well.

https://fuse.wikichip.org/wp-content/uploads/2020/02/isscc2020-prior-comp.png

About 30% of AMD's 50% improvement in perf per Watt with Navi10 over Vega 10 came from N7. The uArch accounted for 60%, and other power optimisations made up another 10%.

People are really overestimating the effect of jumping to N7, seriously.

 

[uzzi38]

1080Ti and 2080Ti are not on the same node. One is 16nm, one is 12nm.

Only difference between 16nm and 12nm is reticle limit afaik. No real efficiency gains past a maturing node.

 

[Hitman928]

Only difference between 16nm and 12nm is reticle limit afaik. No real efficiency gains past a maturing node.

I don't have access to the TSMC 12 nm node but when they announced it the CEO made it sound more like an actual new (half) node.

On the call, TSMC Co-CEO C.C. Wei told analysts that its strategy is "continuously to improve every node in the performance, such as 28-nanometer." He went on to say that TSMC is "continuing to [improve] the 16-nanometers technology." Wei explained that its next revision of the 16nm technology may be worth calling 12nm because it will deliver improved "density, classical density, performance, and power consumption,"

Taiwan Semiconductor Mfg. Co. Ltd. Confirms “12nm” Chip Technology Plans | The Motley Fool

As the competition for more mature chip manufacturing technologies heats up, TSMC isn't standing still.

www.fool.com

 

[DXDiag]

I don't have access to the TSMC 12 nm node but when they announced it the CEO made it sound more like an actual new (half) node.

No, it's just the increased reticle limit for 16nm.

About 30% of AMD's 50% improvement in perf per Watt with Navi10 over Vega 10 came from N7. The uArch accounted for 60%, and other power optimisations made up another 10%.

And you think NVIDIA won't have any of that? Take some 30% from N7, add a minimum of 30% of uArch optimizations (more If NVIDIA revamped the cores as expected), and you have your 70%.

Look, your fallacy is assuming everything else being equal, which is not the case. The cost per gate does not decrease remotely as much when going to 7nm compared to the nodes you are referring to with your examples.

I am willing to discuss this further with you, can you elaborate more?

 

[amrnuke]

Only difference between 16nm and 12nm is reticle limit afaik. No real efficiency gains past a maturing node.

Exactly. Unlike GF 12LP vs Samsung 14LP on which it was licensed, where GF got 10% more density -- TSMC 12 and 16 have no differences in any meaningful respect, except like you said, reticle was widened. I can't say that the masking process wasn't refined and as a result there may be some power and heat savings, but density wise nothing changed.

Against the 1080 Ti, as a result, it's likely that the increased efficiency (17% 2080 Ti over 1080 Ti at 4K) and overall speed (34% faster at 4K) comes from several areas: uarch refinement, 7% more power draw, 64% more die area, GDDR6 RAM instead of GDDR5X, so despite slower clocks it's just a far better performance part because they threw over 50% more transistors at the problem.

I agree - I really don't think anything with the node 16 vs 12 played a major role except that perhaps maturity made it cheaper to make a large die.

 

[AtenRa]

No, it's just the increased reticle limit for 16nm.


And you think NVIDIA won't have any of that? Take some 30% from N7, add a minimum of 30% of uArch optimizations (more If NVIDIA revamped the cores as expected), and you have your 70%.

I am willing to discuss this further with you, can you elaborate more?

He is talking about perf/watt (efficiency) and not performance alone.

Nobody said that NVIDIA will not benefit greatly for 7nm process to increase perf/watt (Efficiency), but asking 70% higher "Raster" performance over the RTX2080Ti is not an easy job now that NVIDIA introduced RTX and Tensor cores in the consumer products.

ps. Also to point out that comparing VEGA 10 to NAVI 10 in Gaming performance is apples to oranges, especially in perf/watt. VEGA mArch wasnt created for gaming, NAVI 10 on the other hand is a evolution over Polaris and not VEGA. But AMD PR took the least capable gaming card (VEGA64) and compared against NAVI 10 to make it look much better in paper

 

[uzzi38]

And you think NVIDIA won't have any of that? Take some 30% from N7, add a minimum of 30% of uArch optimizations (more If NVIDIA revamped the cores as expected), and you have your 70%.

You should probably read what I wrote again. 30% of what AMD gained with RDNA - which was a 50% power efficiency bump - was provided with N7.

That actually means there was a 15% bump in efficiency at the same clocks thanks to the node. Simple maths.

Going from the 1080Ti to the 2080Ti is a 40% bump in perf at 4K at the same TDP with about 17.5% more shaders. So when barely switching node, the improvements from uArch are - no matter how you look at it - below 30%.

Do you honestly think Ampere will not only be a node shift to the most difficult node thus far by a very long shot (see: https://fuse.wikichip.org/news/3320/7nm-boosted-zen-2-capabilities-but-doubled-the-challenges/) but at the same time also such a radical uArch shift with per shader performance improving by a margin significantly larger than Turing did?

 

[DXDiag]

Do you honestly think Ampere will not only be a node shift to the most difficult node thus far by a very long shot (see: https://fuse.wikichip.org/news/3320/7nm-boosted-zen-2-capabilities-but-doubled-the-challenges/) but at the same time also such a radical uArch shift with per shader performance improving by a margin significantly larger than Turing did?

Yes. Turing is 3 years old at this point, it's just an RT evolution of Volta which was introduced 3 years ago. NVIDIA will revamp the uarch.

Take a look at the provided leaked benchmarks, the 7552 core Ampere GPU (@1100MHz) is beating the 4608 core Turing GPU (@1800MHz) by 40% despite using lower clocks, which means this is a ~40% increase in IPC alone. Imagine if the 7552 core GPU operated at something like 1600MHz or more.

 

[CastleBravo]

I think Nvidia could put out a 3080 Ti with 70% more non-RTX performance than the 2080 Ti, but only if it is a 7nm 700+mm^2 monster with an MSRP of $1,500+. I don't see a 70% performance per dollar increase for the top-end card in anything other than ray tracing.