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.

 

[lobz]

Just a very quick and simple estimation:
Taking the average of the die sizes of GTX 680 and GTX 1080, the last GPUs on a new node, you get 304 mm2.

If RTX 3080 is 304 mm^2 in standard TSMC 7nm, that should be around the same number of transistors as a 912 mm^2 die in the previous node (I couldn't find an exact density increase data, should be about 3x).

That's 67% more transistors than RTX 2080. So 75% more powerful seems plausible.

Anyone with better estimations?

3x density is not happening. It's ca. 2x between 12nm (NVIDIA's current node) vs 7nm. That being said, this GPU will be a LARGE monster.
Current flagship (TU102) is 754 mm2. Their compute flagship (Volta, V100) is 815 mm2. I'm not sure what you're talking about.

 

[Adonisds]

3x density is not happening. It's ca. 2x between 12nm (NVIDIA's current node) vs 7nm. That being said, this GPU will be a LARGE monster.
Current flagship (TU102) is 754 mm2. Their compute flagship (Volta, V100) is 815 mm2. I'm not sure what you're talking about.

I was clearly not talking about GA100 but about the RTX 2080. It could also be 75% more powerful.

The RTX 3080 Ti obviously will not be as big as a 2080 Ti after changing to a new node.

Wikichip says 7nm is 3.3X the density of 16nm. And 12nm is 16nm with some small changes. I couldn't find a 16nm to 12nm density change but it should be small

 

[Glo.]

I was clearly not talking about GA100 but about the RTX 2080. It could also be 75% more powerful.

The RTX 3080 Ti obviously will not be as big as a 2080 Ti after changing to a new node.

Wikichip says 7nm is 3.3X the density of 16nm. And 12nm is 16nm with some small changes. I couldn't find a 16nm to 12nm density change but it should be small

You forget that Next gen Nvidia GPUs have to pack more rasterization performance, and more Ray Tracing Performance. What this costs Nvidia? Die size. The die sizes will be equally inflated as Turing's were. Maybe not to the same degree, but I would not expect anything slaller than 600 mm2 from RTX 2080 Ti segment replacement.

P.S. TSMC has missed their power/performance/density targets badly, but was able to make this node Yieldable, compared to Samsung. Don't believe that next gen nodes will not get the same type of hiccups.

 

[Adonisds]

You forget that Next gen Nvidia GPUs have to pack more rasterization performance, and more Ray Tracing Performance. What this costs Nvidia? Die size. The die sizes will be equally inflated as Turing's were. Maybe not to the same degree, but I would not expect anything slaller than 600 mm2 from RTX 2080 Ti segment replacement.

P.S. TSMC has missed their power/performance/density targets badly, but was able to make this node Yieldable, compared to Samsung. Don't believe that next gen nodes will not get the same type of hiccups.

I compared RTX 3080 to RTX 2080. Both have AI, RTX and rasterization parts, so I didn't forget that. It assumes that both will dedicate the same percentage of the die to each part. Reality should be a bit different.

GTX 680 and GTX 1080, the ones that moved to a new node were much smaller than their antecessor and they all only had rasterization, so that's also a fair comparison

My projection is very simple indeed and could be improved a lot, but so far nobody made any valid improvement

 

[Adonisds]

3x density is not happening. It's ca. 2x between 12nm (NVIDIA's current node) vs 7nm. That being said, this GPU will be a LARGE monster.
Current flagship (TU102) is 754 mm2. Their compute flagship (Volta, V100) is 815 mm2. I'm not sure what you're talking about.

I found better data. You're right. Looks like 7nm is only 1.97X as dense. If I understand right they will use 7nm HPC, but there was no HPC lower density version of 16nm or 12nm.

7nm vs 10nm vs 14nm: Fabrication Process - Tech Centurion

Everything about 7nm, 10nm and 14nm Fabrication Process of TSMC, Intel, and Samsung. Transistor Density & Ranking of different technology nodes.

www.techcenturion.com

 

[happy medium]

And dont forget Turing was about 30% faster with the same amount of cores as pascal.

So matching cores for Turing vs Ampere will not work.

Gtx1080 and the rtx2070 super have the same amount of cores and ROP's and frequency. Yet its 30% faster.

 

[lobz]

And dont forget Turing was about 30% faster with the same amount of cores as pascal.

So matching cores for Turing vs Ampere will not work.

Gtx1080 and the rtx2070 super have the same amount of cores and ROP's

But different / more distributed scheduling and such, bigger caches, all of which increase die area, so of course it has higher IPC. Let's see how NVIDIA can clock on 7nm Don't forget, it's easier to go from 1500 to 20-30% higher, than to do the same from 1750-1900.

 

[lobz]

But different / more distributed scheduling and such, bigger caches, all of which increase die area, so of course it has higher IPC. Let's see how NVIDIA can clock on 7nm So far they've done a bang-up job.
However, don't forget, it's easier to go from 1500 to 20-30% higher, than to do the same from 1750-1900.

 

[lobz]

I was clearly not talking about GA100 but about the RTX 2080. It could also be 75% more powerful.

Man, the article says 'The original plan was to outfit the system with Nvidia V100 GPUs' which are 815 mm2 large.

The RTX 3080 Ti obviously will not be as big as a 2080 Ti after changing to a new node.

GP102 471 mm2 went to TU102 754 mm2 on the same process, so you'll never know with NVIDIA

Wikichip says 7nm is 3.3X the density of 16nm. And 12nm is 16nm with some small changes. I couldn't find a 16nm to 12nm density change but it should be small

Those are different processes than what the GPUs use, but you've realized that too in your answer later

 

[happy medium]

But different / more distributed scheduling and such, bigger caches, all of which increase die area, so of course it has higher IPC. Let's see how NVIDIA can clock on 7nm Don't forget, it's easier to go from 1500 to 20-30% higher, than to do the same from 1750-1900.

I'm expecting a 20% clock increase and a 25% per clock boost in performance.
Plus I'm expecting
A 20% reduction in die size.

 

[Glo.]

I compared RTX 3080 to RTX 2080. Both have AI, RTX and rasterization parts, so I didn't forget that. It assumes that both will dedicate the same percentage of the die to each part. Reality should be a bit different.

GTX 680 and GTX 1080, the ones that moved to a new node were much smaller than their antecessor and they all only had rasterization, so that's also a fair comparison

My projection is very simple indeed and could be improved a lot, but so far nobody made any valid improvement

Once again, TSMC has missed power/performance/desity targets. Don't compare one node transition to another.

I'm expecting a 20% clock increase and a 25% per clock boost in performance.

Yeah, and french fries to all of that.

 

[lobz]

I'm expecting a 20% clock increase and a 25% per clock boost in performance.
Plus I'm expecting
A 20% reduction in die size.

All that would mean 50% faster with less than 50% power usage. Not happening.

 

[happy medium]

Once again, TSMC has missed power/performance/desity targets. Don't compare one node transition to another.

Yeah, and french fries to all of that.

They missed power targets for who? Nvidia?
Link me. Yea please link me the proof they missed anything with Nvidia's architecture.

Well dont bother because there is no information about such lies on the internet.

 

[uzzi38]

Once again, TSMC has missed power/performance/desity targets. Don't compare one node transition to another.

Yeah, and french fries to all of that.

TSMC didn't miss perf targets; they were bang on. Please don't tell me you were thinking of Ryzen when you wrote that...

 

[Glo.]

They missed power targets for who? Nvidia?
Link me. Yea please link me the proof they missed anything with Nvidia's architecture.

Oh, so you believe now that Nvidia will magically fix TSMCs 7 nm process?

 

[uzzi38]

All that would mean 50% faster with less than 50% power usage. Not happening.

He said a 20% bump to clocks and 25% bump to perf per clock. That's not 50% perf with half the power, that's 50% perf at the same power.

Whether or not it's doable will be interesting to see. The 30-50% generation on generation performance improvement is totally doable and standard for Nvidia.

 

[Glo.]

TSMC didn't miss perf targets; they were bang on. Please don't tell me you were thinking of Ryzen when you wrote that...

Weren't performance targets higher than what actually we have got? . They missed them just by a little margin. Not as big as power and desnity, but still.

 

[happy medium]

Oh, so you believe now that Nvidia will magically fix TSMCs 7 nm process?

No I'm telling you you have no idea what TSMC has done with Nvidia's architecture.
And your more of a doom and gloom poster for anything Nvidia or for that matter Intel also.
You make up stories,present them as facts and after your proven wrong when the cards hit the shelves, your know where to be found.
That's what Im Saying.

 

[Glo.]

He said a 20% bump to clocks and 25% bump to perf per clock. That's not 50% perf with half the power, that's 50% perf at the same power.

Whether or not it's doable will be interesting to see. The 30-50% generation on generation performance improvement is totally doable and standard for Nvidia.

That requires changing the architecture. It may happen with Ampere, which is HPC only, but Gaming, next gen cards? Im doubtful...

On the other hand, Nvidia has plenty of time to actually put the new arch to gaming cards, so who knows? Definitely will be interesting to see. As I liked Volta and Turing archs, I would love to see where they can push the next gen boudaries, from architectural point of view(ignoring the Ray Tracing BS).

 

[happy medium]

TSMC didn't miss perf targets; they were bang on. Please don't tell me you were thinking of Ryzen when you wrote that...

Yea dont feed him,its a waste of time. But yes you are correct.

 

[uzzi38]

Weren't performance targets higher than what actually we have got? . They missed them just by a little margin. Not as big as power and desnity, but still.

Was 20% perf at same power iirc. Radeon VII demonstrated exactly that - it clocked 200mhz higher and at slightly less power on average (though higher peaks) than Vega 64.

EDIT: Okay, so the above was off memory, checking TPU shows sustained power draw when gaming was like 30W lower, which is quite a bit.

The 20% perf at same power only applies at low clocks, not high clocks (a la Ryzen).

 

[Glo.]

No I'm telling you you have no idea what TSMC has done with Nvidia's architecture.
And your more of a doom and gloom poster for anything Nvidia or for that matter Intel also.
You make up stories,present them as facts and after your proven wrong when the cards hit the shelves, your know where to be found.
That's what Im Saying.

Then show proofs for Your bloody claims, before you will call anyone out. I asked you this before, in another topic. You haven't responded because you haven't got any of them.

So I ask you. Show proofs, that Nvidia will have suddenly better process for their GPUs, and that TSMC will not miss power/density targets for only their GPUs. Show proofs that Ampere is not HPC only. Show proofs that next gen Nvidia cards are actually coming this year, that are not based only on GA100 chip.

Show actual proofs, because when you have been shown proofs or confirmation, that is not happening, you dismissed them, or ignored, just like Muhammad did.

And whats worse, based on your belief that Nvidia will actually deliver next gen GPUs for gaming this year, you were advicing in another thread someone to just wait for them.

You may actually be right, if anyone is patient enough to wait circa 9 months for them, in the best possible scenario.

 

[railven]

No I'm telling you you have no idea what TSMC has done with Nvidia's architecture.
And your more of a doom and gloom poster for anything Nvidia or for that matter Intel also.
You make up stories,present them as facts and after your proven wrong when the cards hit the shelves, your know where to be found.
That's what Im Saying.

Nailed it! Glad someone else is tired of it. I'm still waiting for the magic drivers that were suppose to make Vega ROFLOL stomp Pascal.

 

[uzzi38]

That requires changing the architecture. It may happen with Ampere, which is HPC only, but Gaming, next gen cards? Im doubtful...

On the other hand, Nvidia has plenty of time to actually put the new arch to gaming cards, so who knows? Definitely will be interesting to see. As I liked Volta and Turing archs, I would love to see where they can push the next gen boudaries, from architectural point of view(ignoring the Ray Tracing BS).

You think the delay will happen WITHOUT Nvidia tweaking the uArch?

Man, not even I'm that optimistic haha.

 

[Glo.]

Was 20% perf at same power iirc. Radeon VII demonstrated exactly that - it clocked 200mhz higher and at slightly less power on average (though higher peaks) than Vega 64.

EDIT: Okay, so the above was off memory, checking TPU shows sustained power draw when gaming was like 30W lower, which is quite a bit.

The 20% perf at same power only applies at low clocks, not high clocks (a la Ryzen).

Fair enough. Then yes, TSMC did not missed their Performance targets, contrary to density and power targets.

 

[Adonisds]

And dont forget Turing was about 30% faster with the same amount of cores as pascal.

So matching cores for Turing vs Ampere will not work.

Gtx1080 and the rtx2070 super have the same amount of cores and ROP's and frequency. Yet its 30% faster.
View attachment 16617

As lobz said it's the same number of cores but there's much more transistors per core, so it didn't come for free.

Man, the article says 'The original plan was to outfit the system with Nvidia V100 GPUs' which are 815 mm2 large.

GP102 471 mm2 went to TU102 754 mm2 on the same process, so you'll never know with NVIDIA

Those are different processes than what the GPUs use, but you've realized that too in your answer later

Pascal to Turing was a transition using almost the same node. 12nm is 16nm with small changes. So that's why the die size increased just like in the Geforce 600 to 900 transition. It's not happening this time, it will be smaller.