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]

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?

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.

My issue is that TSMC claims that AMD and Apple are equal sized customers for 7nm, BEFORE the move to 5nm. At the same time we have claims that Apple is bigger than both AMD and Nvidia combined so there's no issue with Nvidia finding space to fab their chips at TSMC once Apple moves to 5nm later this year. I'm saying that this is wrong based on TSMC's direct statement of present usage.

 

[insertcarehere]

My issue is that TSMC claims that AMD and Apple are equal sized customers for 7nm, BEFORE the move to 5nm.

Source? All similar news I can find relating to TSMC's 7nm orders comes from Taiwanese Media speculation, not TSMC themselves.

And even if that source is correct, those figures quoted still don't account for ~30% capacity open to split between Apple & Nvidia in 'H2 2020, even if Apple still takes ~20k wafers/month (unlikely given 5nm ramp-up) that leaves a healthy share for Nvidia to get significant 7nm product out by year-end. It takes a lot less wafers to get the same volume with 7nm vs 12nm.

 

[Stuka87]

Source? All similar news I can find relating to TSMC's 7nm orders comes from Taiwanese Media speculation, not TSMC themselves.

And even if that source is correct, those figures quoted still don't account for ~30% capacity open in 'H2 2020, even if Apple still takes ~20k wafers/month (unlikely given 5nm ramp-up) that leaves a healthy share for Nvidia to get significant 7nm product out by year-end. It takes a lot less wafers to get the same volume with 7nm vs 12nm.

There won't be 30% available though. The article states that AMD bought up a bunch of wafers that Apple would have taken. With AMD being a Tier 1 customer, they are going to get their choice of wafers ahead of nVidia. nVidia may sell more GPU's than AMD, but when it comes to total chips manufactured, AMD dwarfs nVidia. Especially since a big chunk of that capacity that AMD grabbed is for the new Xbox/PS consoles. Plus Ryzen still selling like bonkers, and the new big core count chips coming out.

So, Ampere will most likely be on TSMC, but that one chip. The consumer grade chips, once they come out, will most likely be samsung.

 

[Glo.]

So, Ampere will most likely be on TSMC, but that one chip. The consumer grade chips, once they come out, will most likely be samsung.

That is assuming Samsung can deliver their process.

For which it appears at first glance is either broken from performance/watt point of view, or yield rate.

Or both, at the same time.

 

[Stuka87]

That is assuming Samsung can deliver their process.

For which it appears at first glance is either broken from performance/watt point of view, or yield rate.

Or both, at the same time.

True, I am assuming that they may have it resolved by later this year. But if not, and nVidia had developed their chips to use it, thats going to hurt, as they can't just move the chips over to TSMC, there are enough differences that it requires some work.

 

[maddie]

Source? All similar news I can find relating to TSMC's 7nm orders comes from Taiwanese Media speculation, not TSMC themselves.

And even if that source is correct, those figures quoted still don't account for ~30% capacity open to split between Apple & Nvidia in 'H2 2020, even if Apple still takes ~20k wafers/month (unlikely given 5nm ramp-up) that leaves a healthy share for Nvidia to get significant 7nm product out by year-end. It takes a lot less wafers to get the same volume with 7nm vs 12nm.

Generally die sizes do not fall with GPUs, but the increased transistor budget is used for increasing performance. It's not like they're just porting the RTX 2xxx series to 7nm. I would assume wafer counts will not fall a lot if at all.

 

[insertcarehere]

Generally die sizes do not fall with GPUs, but the increased transistor budget is used for increasing performance. It's not like they're just porting the RTX 2xxx series to 7nm. I would assume wafer counts will not fall a lot if at all.

You'd have a point if TU102, TU104 and TU106 weren't already significant increases in size over their Pascal counterparts. It doesn't take an electrical engineer to guess that trying to make a GP102-sized chip on 7nm will be extremely difficult. I would expect Ampere to revert to the historic norms in terms of Nvidia die sizes (500-600mm^2 for Ampere002, 300-400mm^2 for Ampere004 etc). Which would account for a good amount extra chips/wafer.

Not to mention that Nvidia are in no rush to transition their lower-end products to TSMC 7nm any time soon. The non-RTX Turings are competitive enough that they could stick around on 12nm for a while, and it also won't be surprising if the non-RTX Amperes go to Samsung 7nm given that JHH has ordered volume from them this year.

 

[Mopetar]

NVidia probably could get away with a Turing die shrink if they absolutely needed new cards to compete. I think they've got so much room with potential price cuts that they don't even need to rush to get to 7nm. Even if NVidia isn't going to have anything out until late this year or early next, it won't matter because AMD won't have anything outside of a high-end Navi card. Even if it were able to match (or even overtake) the 2080 Ti, NVidia isn't going to lose customers to AMD over it.

 

[jpiniero]

I think they would like to get new cards out this year, to continue to push RT.

I have no doubts they can get the wafers. My question is mainly do they have the designs completed at TSMC. A Turing shrink and higher clocks would be fine, and then just skip Ampere for gaming like they did Volta.

 

[RetroZombie]

It's interesting that amd skipped tsmc 16nm and nvidia apparently skipped tsmc 7nm, I wonder why since both companies have partnering with them for years.
Some burnen with them?

 

[Ottonomous]

What's the possibility of nvidia selling the '2nd gen RTX' on top of the first gen to create a new pricing scheme?

 

[Ajay]

Man,

I think they would like to get new cards out this year, to continue to push RT.

I have no doubts they can get the wafers. My question is mainly do they have the designs completed at TSMC. A Turing shrink and higher clocks would be fine, and then just skip Ampere for gaming like they did Volta.

Yeah, if NV had to shift from Samsung to TSMC for at least part of their 7nm portfolio, that must have set them back quite a bit.

 

[Ajay]

It's interesting that amd skipped tsmc 16nm and nvidia apparently skipped tsmc 7nm, I wonder why since both companies have partnering with them for years.
Some burnen with them?

NV moved, supposedly, for a great deal from Samsung - who where keen to land a marquee semiconductor brand.

 

[Muhammed]

That is assuming Samsung can deliver their process.

But if not, and nVidia had developed their chips to use it, thats going to hurt, as they can't just move the chips over to TSMC,

WRONG! God, you two are just repeating each other at this point. NVIDIA is using TSMC 7nm for ALL of it's next gen lineup, all of this crap about Samsung is completely wrong.

The information comes straight from NVIDIA's CEO who stated during the media Q/A session that they will primarily be selecting TSMC for the majority of their next-generation 7nm GPU orders with Samsung only receiving a smaller number of orders.

NVIDIA Next-Generation 7nm GPU: TSMC To Get The Bulk of Orders, Samsung's EUV Has A Smaller Role

During GTC 2019, NVIDIA's CEO responded to the press that the majority orders of their next-generation 7nm GPU will be handled by TSMC

wccftech.com

 

[Ajay]

WRONG! God, you two are just repeating each other at this point. NVIDIA is using TSMC 7nm for ALL of it's next gen lineup, all of this crap about Samsung is completely wrong.

NVIDIA Next-Generation 7nm GPU: TSMC To Get The Bulk of Orders, Samsung's EUV Has A Smaller Role

During GTC 2019, NVIDIA's CEO responded to the press that the majority orders of their next-generation 7nm GPU will be handled by TSMC

wccftech.com

Okay, that was interesting. I wonder how the Samsung order got so blown out of proportion. Now we’re stuck hanging around waiting to see what NV has in store for gamers.

 

[Stuka87]

Okay, that was interesting. I wonder how the Samsung order got so blown out of proportion. Now we’re stuck hanging around waiting to see what NV has in store for gamers.

Samsung and nVidia both said on several occasions over the summer that Samsung would be getting a "substantial" number of orders for 7nm chips.

Its only in the last few weeks that reports of nVidia using TSMC, and not Samsung have come out.

WRONG! God, you two are just repeating each other at this point. NVIDIA is using TSMC 7nm for ALL of it's next gen lineup, all of this crap about Samsung is completely wrong.

Ok, you are now contradicting the very report you just posted. Where is states, and I quote: "they will primarily be selecting TSMC for the majority of their next-generation 7nm GPU orders with Samsung only receiving a smaller number of orders". Which means, not all orders will be going to TSMC.

Either way, which ever one they use, its still unlikely that consumer model chips will be released in the first half of the year. GA100 will most likely be unveiled as a Tesla card, this is pretty much expected.

 

[jpiniero]

Okay, that was interesting. I wonder how the Samsung order got so blown out of proportion. Now we’re stuck hanging around waiting to see what NV has in store for gamers.

They were going to use Samsung mostly. Now they aren't because the yields are unsatisfactory.

 

[JasonLD]

They were going to use Samsung mostly. Now they aren't because the yields are unsatisfactory.

Well, it is unknown whether Nvidia was going to use Samsung for their GPUs or their consumer GPUs were always planned for TSMC in the first place. I don't expect Nvidia ever planned to use Samsung for all of their GPU line up since Samsung is relatively inexperienced in making big chips compared to TSMC.

 

[Glo.]

WRONG! God, you two are just repeating each other at this point. NVIDIA is using TSMC 7nm for ALL of it's next gen lineup, all of this crap about Samsung is completely wrong.

NVIDIA Next-Generation 7nm GPU: TSMC To Get The Bulk of Orders, Samsung's EUV Has A Smaller Role

During GTC 2019, NVIDIA's CEO responded to the press that the majority orders of their next-generation 7nm GPU will be handled by TSMC

wccftech.com

It says: Majority. And If I understand English language well enough, Majority means: "not all of" 😉.

So this means directly: Nvidia will use TSMC for major part of their orders of next gen products. So what else there is that has 7 nm process, hmmm? Which other vendor has it, hmmm? 😉

You have direct confirmation of what we are saying. Nvidia went to samsung first for their products, but considering the Yield and performance is unsatisfactory, they went back to TSMC.

And porting of ONE Product to another, non-compatible process takes time.

I will repeat this. 7 nm Products from Nvidia we will see in 2020 is only HPC GA100 Chip. Gaming cards are 7 nm, and Late 2020/Early 2021 products. And Gaming cards are not Ampere architecture. it will be completely different name.

 

[Glo.]

Either way, which ever one they use, its still unlikely that consumer model chips will be released in the first half of the year. GA100 will most likely be unveiled as a Tesla card, this is pretty much expected.

Considering the only recent switch from Samsung to TSMC by Nvidia I would raise a question.

Have gaming cards even been taped out, yet? 🙂

 

[maddie]

Considering the only recent switch from Samsung to TSMC by Nvidia I would raise a question.

Have gaming cards even been taped out, yet? 🙂

Maybe they covered themselves by having designs for both fabs until things became clearer. They certainly have enough resources to do this as most of the R&D before committing, would have gone into the architectural layouts of the design. Cache , shader layout, RTX, etc. That's the only way they could shift so fast as his statement seems almost as if they're very late in deciding.

 

[Glo.]

Maybe they covered themselves by having designs for both fabs until things became clearer. They certainly have enough resources to do this as most of the R&D before committing, would have gone into the architectural layouts of the design. Cache , shader layout, RTX, etc. That's the only way they could shift so fast as his statement seems almost as if they're very late in deciding.

Well, lets hope so, that is the case, really.

 

[Ajay]

Considering the only recent switch from Samsung to TSMC by Nvidia I would raise a question.

Have gaming cards even been taped out, yet? 🙂

The quoted articles say nothing about a 'switch'. So, now the question become, was Samsung ever going to be producing a significant portion of NV GPUs. Seems like the answer is not, unless there is countervailing evidence.

 

[Ajay]

Samsung and nVidia both said on several occasions over the summer that Samsung would be getting a "substantial" number of orders for 7nm chips.

Its only in the last few weeks that reports of nVidia using TSMC, and not Samsung have come out.

Substantial orders of chips doesn't necessarily mean GPUs. Something is rotten in Denmark at the moment, IMO.

 

[AtenRa]

TSMCs 7nm volume production is increasing in 2020, its not stagnant. Add that Apple will move to 5nm in 2020 and there will be plenty of wafer volume for NVIDIA even though AMD will increase its allocation in H2 2020.

Also, 7nm+ EUV should be cheaper than 7nm Triple patterning was in 2019 due to less design rules, less masks, less production steps etc etc

https://semiwiki.com/wp-content/uploads/2018/01/Scott20Jones20ISS20201820Presentation-1.pdf