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.

 

[maddie]

Apple is moving to 5 nm, and several other mobile chip makers are too. That will free up a lot of volume.

JHH said they are mostly using TSMC.

You see enough capacity for both Nvidia & AMD with AMD going to become the top volume client on 7nm? With Nvidia having 70-80% of the discrete GPU market and GPUs being much bigger dies than AMDs CPU chiplet?

 

[jpiniero]

You see enough capacity for both Nvidia & AMD with AMD going to become the top volume client on 7nm? With Nvidia having 70-80% of the discrete GPU market and GPUs being much bigger dies than AMDs CPU chiplet?

Uh yes? Apple is bigger than both combined, easy.

 

[Stuka87]

Samsung EUV 7nm . This process will be more efficient than even 7nm TSMC.
I'm guessing the 3080 will be at least 20% faster than a gtx2080ti.
Save my posts if you like.

Historically, this is not been the case. The 2080 sure isn't anywhere close to being 20% faster than a 1080Ti. Typically the top non-Ti card matches the previous Ti card. The RTX 2080 is a few percent faster than a 1080Ti, the 2080S is faster yet.

I think it comes down to if nVidia wants to make rasterized games faster, or RTX games faster. Personally, I am expecting them to spend most of the newly available die space on RTX hardware. If they want ray tracing to take off, they have to make it playable, and currently it is not (Spending $1000 to play at 1080P is not acceptable). If they do this, the 3080 will most likely just match the 2080Ti, and not be any faster for non-RTX games.

 

[IntelUser2000]

Historically, this is not been the case. The 2080 sure isn't anywhere close to being 20% faster than a 1080Ti. Typically the top non-Ti card matches the previous Ti card. The RTX 2080 is a few percent faster than a 1080Ti, the 2080S is faster yet.

You sure they can't do both?

The RTX series not only added RT and Tensor cores, it was on the same process. Their next generation is 7nm.

The GTX 1080 beat the 980 Ti. Now its very possible we'll see a staggered launch with RTX 3080 coming say middle of this year, and RTX 3080 Ti end of this year.

 

[Glo.]

You sure they can't do both?

The RTX series not only added RT and Tensor cores, it was on the same process. Their next generation is 7nm.

The GTX 1080 beat the 980 Ti. Now its very possible we'll see a staggered launch with RTX 3080 coming say middle of this year, and RTX 3080 Ti end of this year.

I will repeat this once again. Because Samsung failed to deliver 7 nm process, Nvidia hasn't even taped out their consumer, gaming architecture.

Gaming GPUs from Nvidia will not happen before late 2020, or early 2021. Period. Only 7 nm GPU we will see this year is Ampere, GA100, for HPC, that replaces Volta.

Period.

 

[railven]

You sure they can't do both?

The RTX series not only added RT and Tensor cores, it was on the same process. Their next generation is 7nm.

The GTX 1080 beat the 980 Ti. Now its very possible we'll see a staggered launch with RTX 3080 coming say middle of this year, and RTX 3080 Ti end of this year.

Things can get really ugly if NV's move to 7nm is anything like it's move to 40nm. AMD raised the price on their Polaris replacement just because it could compete with NV's current products. They matched them perf / price. If NV can deliver RTX 2070/5700 XT with their THIRD chip who's to stop them from placing it right at the same price point?

AMD's current top tier GPU is competing with NV's cut down second GPU. One has a node advantage. People are already making excuses for AMD raising prices just like the the otherside did when NV raised prices in 2012.

EDIT: Even the AMD users are starting to realize they've gotten fleeced.

https://www.reddit.com/r/Amd/comments/eo26de

 

[GodisanAtheist]

Historically, this is not been the case. The 2080 sure isn't anywhere close to being 20% faster than a 1080Ti. Typically the top non-Ti card matches the previous Ti card. The RTX 2080 is a few percent faster than a 1080Ti, the 2080S is faster yet.

I think it comes down to if nVidia wants to make rasterized games faster, or RTX games faster. Personally, I am expecting them to spend most of the newly available die space on RTX hardware. If they want ray tracing to take off, they have to make it playable, and currently it is not (Spending $1000 to play at 1080P is not acceptable). If they do this, the 3080 will most likely just match the 2080Ti, and not be any faster for non-RTX games.

The Reference GTX 1080 was ~ 30% faster than the reference 980ti, and roughly 15-20% faster than the AIB 980ti's. The 980ti was no tiny die either, thing was a massive 601mm2 chip.

It can definitely be done with a die shrink, even with these massive RTX dies. NV needs to give the 9xx and 10xx series owners a real reason to upgrade in their price category. Whether they will or not, however, is another story.

 

[IntelUser2000]

@railven Ahahhah! One thing I really got out of Ethereum mining is I now have a good gaming card as well. The RX 470/570s still rock.

 

[linkgoron]

Things can get really ugly if NV's move to 7nm is anything like it's move to 40nm. AMD raised the price on their Polaris replacement just because it could compete with NV's current products. They matched them perf / price. If NV can deliver RTX 2070/5700 XT with their THIRD chip who's to stop them from placing it right at the same price point?

While 5700XT/5700 are indeed priced too high for a Polaris replacement, it's untrue that AMD "matched" Nvidia with the 5700 cards. First, they made Nvidia price-cut Turing (The 2070 essentially went down by $100 with the 2060S launch).

Per Anandtech's review:

The GeForce RTX 2060 Super is all but an RTX 2070 in name and in price, delivering virtually identical performance for $100 less than the original RTX 2070. And the GeForce RTX 2070 Super, while not quite a facsimile of the RTX 2080, delivers much of those gains, offering 96% of the RTX 2080’s performance for 71% of the price – or nearly some $200 cheaper than what that level of performance cost just last month.

Second, (even after the price cuts) according to TPU, the 5700XT easily beats the price/perf of the 2060S, and the 5700 easily beats the price/perf of the 2060, both by more than 10% than their respective competitor.

The numbers are from the latest 5500XT review (As they have the latest numbers):
1080p numbers:


1440p numbers:

 

[Muhammed]

Where is the available wafer volume at TSMC for Nvidia to keep their market share?

Ampere is 7nm+
NVIDIA said they are using TSMC 7nm, Samsung is not on the table for them.

 

[Muhammed]

Because Samsung hasn't solved their process, yet, and at TSMC they will not have enough 7 nm capacity to move enough GPUs for consumer AND HPC market, especially considering the die size of GA100.

Nope, Ampere is 7nm+. And Ampere is for all sectors, HPC, gamers, AI .. etc.
NVIDIA said they are using TSMC 7nm, Samsung is not on the table for them.

Also Ampere taped out early last year.

https://twitter.com/x/status/1112053410860290049

 

[Glo.]

Nope, Ampere is 7nm+. And Ampere is for all sectors, HPC, gamers, AI .. etc.
NVIDIA said they are using TSMC 7nm, Samsung is not on the table for them.

Also Ampere taped out early last year.

https://twitter.com/x/status/1112053410860290049

Yes, they taped out HPC chips.

But not gaming chips. Period. Show me where are gaming device IDs. GA102, GA104, GA106.

There is only ONE GA chip:

https://twitter.com/x/status/1213535242650173440

GA100. There is nothing else. GA100 is HPC chip, and Ampere is ONLY for HPC.

Gaming cards use completely different architecture, and are not in plans for closest future. Which is the reason why Nvidia is readying RTX 2080 Ti Super with 4608 ALUs, full 384 Bits and 16 Gbps GDDR6.

Period.

Also TSMC is on the table, BECAUSE SS's process failed, and SS couldn't deliver the process. Which is the reason why Nvidia is so late to the party of next gen nodes.

 

[Muhammed]

GA100. There is nothing else. GA100 is HPC chip, and Ampere is ONLY for HPC.

Gaming cards use completely different architecture, and are not in plans for closest future. Which is the reason why Nvidia is readying RTX 2080 Ti Super with 4608 ALUs, full 384 Bits and 16 Gbps GDDR6.

Period.

Also TSMC is on the table, BECAUSE SS's process failed, and SS couldn't deliver the process. Which is the reason why Nvidia is so late to the party of next gen nodes.

Your post couldn't be more wrong.

Again, NVIDIA announced next gen cards are using TSMC 7nm+.
Again Ampere taped out in full, NVIDIA GPUs don't leak out in drivers like AMD GPUs do. Show me where Turing or Pascal leaked out in drivers?

Comes mid 2020 you will be hiding your face in shame.

 

[Ottonomous]

Your post couldn't be more wrong.

Again, NVIDIA announced next gen cards are using TSMC 7nm+.
Again Ampere taped out in full, NVIDIA GPUs don't leak out in drivers like AMD GPUs do. Show me where Turing or Pascal leaked out in drivers?

Comes mid 2020 you will be hiding your face in shame.

You mean how Volta was leaked in Gameworks slides and everyone banged on about it's gaming prowess?

Edit: inb4 Volta was a great gaming chip, I still stand correct mister

 

[Ottonomous]

The really weird rumour is how Hopper is associated with MCM GPU. Maybe that's another or the HPC uarch they're touting, and the supposed Aerial is the gaming part.

Would be nice because she popuralized machine-independent languages, so here's guessing they've found solutions to MCM/Dual-GPUs performance issues. But unfortunately like RTX another expensive route to circumvent Moore's law.

 

[railven]

While 5700XT/5700 are indeed priced too high for a Polaris replacement, it's untrue that AMD "matched" Nvidia with the 5700 cards. First, they made Nvidia price-cut Turing (The 2070 essentially went down by $100 with the 2060S launch).

Per Anandtech's review:


Second, (even after the price cuts) according to TPU, the 5700XT easily beats the price/perf of the 2060S, and the 5700 easily beats the price/perf of the 2060, both by more than 10% than their respective competitor.

The numbers are from the latest 5500XT review (As they have the latest numbers):
1080p numbers:
View attachment 15680

1440p numbers:
View attachment 15681

I think you're missing my point. To me NV will never be a perf / price choice. AMD was always was that choice.

The last time AMD priced their chip to compete directly on perf / price with NV's current product stack AMD got wrecked.

AMD went from competing with their top chip against a cut down version of NV's top chip:
HD 5870 vs GTX 470 (cut down GF100)
HD 6970 vs GTX 570 (cut down GF110)

To AMD competing with NV's second top chip in their family
HD 7970 vs GTX 680 (full GK104)
R9 290X vs GTX 980 (full GM104)

To AMD now competing with NV's cut down second chip!
5700 XT vs RTX 2070 Super (cut down TP104)

If whatever NV has stored for their 7nm product line is any good, we can see GWhatever 106 in the >$300 product brackets.

I don't put it past NV to do this. They've shown time and time again that they'll raise prices and we consumers cheer them on because we ain't got nothing else to really choose from.
$300ish 5700 XT/RTX 2060 Super'ish performance from RDNA 1.0 cards or from NV's possible GWhatever 106 chip. And if AMD doesn't clean up their driver issues, NV isn't gonna have an issue if they can deliver.

But I admit, this is my "doom and gloom" outlook. Me believing NV cares, would see their next family of cards bring prices down a little, but even that feels like a pipe dream.

 

[Stuka87]

Nope, Ampere is 7nm+. And Ampere is for all sectors, HPC, gamers, AI .. etc.
NVIDIA said they are using TSMC 7nm, Samsung is not on the table for them.

Also Ampere taped out early last year.

https://twitter.com/x/status/1112053410860290049

1: That "source" was obviously wrong, as it doesn't take a year to release a chip after tape out.

2: Yes, Ampere is TSMC 7nm+, but its GA100. Big chip for Tesla cards, not gaming.

There have been zero leaks of gaming cards. This same thing happened when Volta was getting close to release, and you were wrong then too as Volta was only an HPC card (And no, a Titan V is not a gaming card).

 

[AtenRa]

I believe NVIDIA will heavily invest in RayTracing again ,

So personally im expecting only 30% higher FP32 performance in games but 2x RT performance over Turing architecture.

Turing chips are way too big and even with the 7nm+ density, porting the TU102 (RTX2080Ti) to TSMCs 7nm+ will still be a huge chip at around 400-430mm2. Add more transistors to RT-Tensor Cores and next RTX3080Ti could be close to 450-500mm2.
You can clearly see that FP32 performance for rasterized gaming will not have the necessary transistor budget for more than 30% higher performance IF they invest more in to RT-Tensor Cores.

 

[exquisitechar]

I think you're missing my point. To me NV will never be a perf / price choice. AMD was always was that choice.

The last time AMD priced their chip to compete directly on perf / price with NV's current product stack AMD got wrecked.

AMD went from competing with their top chip against a cut down version of NV's top chip:
HD 5870 vs GTX 470 (cut down GF100)
HD 6970 vs GTX 570 (cut down GF110)

To AMD competing with NV's second top chip in their family
HD 7970 vs GTX 680 (full GK104)
R9 290X vs GTX 980 (full GM104)

To AMD now competing with NV's cut down second chip!
5700 XT vs RTX 2070 Super (cut down TP104)

If whatever NV has stored for their 7nm product line is any good, we can see GWhatever 106 in the >$300 product brackets.

I don't put it past NV to do this. They've shown time and time again that they'll raise prices and we consumers cheer them on because we ain't got nothing else to really choose from.
$300ish 5700 XT/RTX 2060 Super'ish performance from RDNA 1.0 cards or from NV's possible GWhatever 106 chip. And if AMD doesn't clean up their driver issues, NV isn't gonna have an issue if they can deliver.

But I admit, this is my "doom and gloom" outlook. Me believing NV cares, would see their next family of cards bring prices down a little, but even that feels like a pipe dream.

Strange take, in most of your comparisons you are comparing a smaller AMD “top” chip to a bigger NV chip.

AMD will give you what to choose from, but they won’t be your “perf/price” choice. They’ll make you pay for it. You will see this when the biggest RDNA2 card handily beats the 2080ti, but costs more than it too. And yeah, the GPU market is going to be ugly, trust me. Next gen Nvidia gaming GPUs will be no better in terms of pricing.

 

[maddie]

Uh yes? Apple is bigger than both combined, easy.

Using this translation from chiakokua we get some relevant info and questions.

TSMC's 7nm production capacity is fully booked. Relief may only come when Apple migrates to 5nm in 2H'2020. TSMC's 7nm capacity will increase to 140,000 wpm in 2H'2020. By order proportion, the ranking of customers using 7nm will be re-shuffled. AMD's orders are set to double, replacing Apple as the largest customer [for 7nm]. Huawei's HiSilicon and Qualcomm are similar by order proportion.

TSMC's 7nm production capacity continues to rise. The industry expects monthly capacity to reach 110,000 wafers in 1H'2020. The top 5 customers by order proportion are: Apple, HiSilicon, Qualcomm, AMD, and Mediatek. Except for Mediatek, order share is split at roughly 20% each, depending on seasonality. Mediatek's share is around 13%.

However, with 7nm capacity rising to 140,000 wpm in 2H'2020, and the largest customer Apple migrating to 5nm with the A14 processor, customer ranking by 7nm orders will be re-shuffled. In one fell swoop, AMD booked capacity for 30,000 wafers, accounting for 21% of total capacity. HiSilicon and Qualcomm's orders are similar, at 17-18%. Mediatek's share also rose to 14%.

At present, Samsung's 7nm production capacity is roughly 150,000 wpm. It is also actively increasing 7nm capacity. According to industry rumors, Samsung plans to quadruple capacity in 2020. Nvidia and Qualcomm's next-generation products may be produced using Samsung's 7nm EUV process, but details remain to be seen.

1) Presently, top 4 customers, Apple and AMD included, have roughly an equal share of 20% each of roughly 80-100K wafers. So 16-20 K each.

How then can you say that Apple is bigger than AMD and Nvidia in wafer volume combined?

2)TSMC going to 140K wafers in 2H2020 with AMD share rising to ~ 30K wafers.

Isn't this alone going to be bigger than Apple's present wafer use?

3)I don't know if this is accurate, but it has Samsung at 150K wafers at present, going to 600K by 2H2020.

Is Samsung really 150% presently and going to 400% of TSMC in 2H2020, or is the decimal misplaced and we have to divide by 10, 15K > 60K in reality?
How are they sourcing the EUV machines to increase wafer output by 450K this year?

I would also assume that Nvidia uses more wafers than even the present AMD.

All of this tells me that there's no way Nvidia can use TSMC this year to fab consumer GPUs and maintain sales volume if using TSMC.

 

[DrMrLordX]

A few weeks after the big Navi release or the June /July time frame.

You sure that's gonna be Ampere?

 

[jpiniero]

1) Presently, top 4 customers, Apple and AMD included, have roughly an equal share of 20% each of roughly 80-100K wafers. So 16-20 K each.

How then can you say that Apple is bigger than AMD and Nvidia in wafer volume combined?

2)TSMC going to 140K wafers in 2H2020 with AMD share rising to ~ 30K wafers.

Isn't this alone going to be bigger than Apple's present wafer use?

Like I said, they are moving the iPhone production to 5 nm. Even for Apple, it takes months to assemble enough phones, let alone produce the chips themselves. So their lead times are long. Apple usually has TSMC start in March, but this year it might be April or May.

Apple might be at similar volume to AMD with just iPad.

 

[maddie]

Like I said, they are moving the iPhone production to 5 nm. Even for Apple, it takes months to assemble enough phones, let alone produce the chips themselves. So their lead times are long. Apple usually has TSMC start in March, but this year it might be April or May.

Apple might be at similar volume to AMD with just iPad.

When the info was presented, TSMC themselves said that Apple and AMD are equal sized 7nm customers. After they move to 5nm, AMD will become the lead customer on 7nm.

What exactly is unclear about that statement and how can you then make the claim about the iPads?

 

[Glo.]

I think you're missing my point. To me NV will never be a perf / price choice. AMD was always was that choice.

Intel was also the giant that couldn't be possibly taken out.

And yet, their own incompetence has backfired badly for them. What makes you believe Nvidia cannot **** up?

When the info was presented, TSMC themselves said that Apple and AMD are equal sized 7nm customers. After they move to 5nm, AMD will become the lead customer on 7nm.

What exactly is unclear about that statement and how can you then make the claim about the iPads?

Apple will still make 7 nm chips even after they move to 5 nm. Why? They still make "Cheaper" iPhones, iPads, than the bleeding edge ones.

 

[jpiniero]

Apple will still make 7 nm chips even after they move to 5 nm. Why? They still make "Cheaper" iPhones, iPads, than the bleeding edge ones.

True, but I think those are mostly excess inventory and/or quality dud chips that they've stockpiled over time.