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.

 

[gdansk]

GA104 taped out recently, GA103 (probably) cancelled, GA102 taped out on Samsung 10/8nnm in half of March

Twitter - kopite7kimi

//fixed - GA102 taped out in half of March not May

Interesting. Anyone hazard a guess if GA100 will show up in RTX/Titan cards or is that for Tesla/Quadro only?
I'd guess GA102 -> RTX 3080 Ti but maybe they'll bring GA100 as a Titan card later. It's hard to guess because they did release both GV100 and TU102 as Titan cards.

 

[Stuka87]

Interesting. Anyone hazard a guess if GA100 will show up in RTX/Titan cards or is that for Tesla/Quadro only?
I'd guess GA102 -> RTX 3080 Ti but maybe they'll bring GA100 as a Titan card later. It's hard to guess because they did release both GV100 and TU102 as Titan cards.

GA100 will never show up in a consumer card. It is tailored for a specific use case, and doesn't have any hardware ray tracing functionality.

 

[uzzi38]

Interesting. Anyone hazard a guess if GA100 will show up in RTX/Titan cards or is that for Tesla/Quadro only?
I'd guess GA102 -> RTX 3080 Ti but maybe they'll bring GA100 as a Titan card later. It's hard to guess because they did release both GV100 and TU102 as Titan cards.

GA100 has no display engine nor NVENC, so no.

 

[Ajay]

GA104 taped out recently, GA103 (probably) cancelled, GA102 taped out on Samsung 10/8nnm in half of March

Twitter - kopite7kimi

//fixed - GA102 taped out in half of March not May

So, unless I am freaking blind, I didn't see any mention of Samsung being the target fab. (Edit, should have clicked on second link - still, what in the freaking hell is NV doing?!).
At this point in time, it would only make sense to target 7nm at either SS or TSMC.

Also, I think tapeout to retail availability is typically around a year, give or take. I suppose NV anticipated this and has added resources to reduced the TTM. Anyway, really bad news for NV if true.

 

[uzzi38]

So, unless I am freaking blind, I didn't see any mention of Samsung being the target fab.
At this point in time, it would only make sense to target 7nm at either SS or TSMC.

Also, I think tapeout to retail availability is typically around a year, give or take. I suppose NV anticipated this and has added resources to reduced the TTM. Anyway, freaking bad news for NV if true.

https://twitter.com/x/status/1237687606869188608

 

[gdansk]

GA100 will never show up in a consumer card. It is tailored for a specific use case, and doesn't have any hardware ray tracing functionality.

GA100 has no display engine nor NVENC, so no.

Ah. So then like TU102, GA102 will probably be cut down for the top of the line 3000 series card. For shame

 

[Ajay]

https://twitter.com/x/status/1237687606869188608

Yeah, realized I didn't click on the second link. Still, with tapeout occurring THIS year, why not target 7nm? I don't get it.

 

[uzzi38]

Yeah, realized I didn't click on the second link. Still, with tapeout occurring THIS year, why not target 7nm? I don't get it.

TSMC N7 may have not been targetted due to wafer constraints. After trying to use 7LPP, which was a flop due to poor yields, Nvidia had two choices. Eat up wafer supply that was allocated for GA100 or switch to Samsung's 8LPP instead.

There aren't enough N7 wafers to go around, getting ahold of enough supply for all their consumer GPUs is difficult. That being said I wouldn't be surprised if they use N7 for at least one of their consumer dies, even if it's a retape they spin after the initial lineup rolls out.

I suspect this is also why the rumour that Nvidia are buying N5 wafers early has come about. They don't want to be in this situation again.

I can say seperately I'm confident that Samsung ran into major issues with 7LPP, because I know that it had a knock on effect with their 5nm later down the line. The issue was likely tackled afaik... but you don't just fix an issue like this without knock on effects. And there were definitely knock on effects.

 

[Krteq]

Yeah, realized I didn't click on the second link. Still, with tapeout occurring THIS year, why not target 7nm? I don't get it.

NV tried to manipulate TSMC for better prices, so they taped out some chips at Samsung... but their risk didn't ended well... other companies took their opportunities and split up a rest of 7nm capacity from leaving Apple. So NV had to manufacture some chips at Samsung because there are no capacity left at TSMC

 

[Ajay]

TSMC N7 may have not been targetted due to wafer constraints. After trying to use 7LPP, which was a flop due to poor yields, Nvidia had two choices. Eat up wafer supply that was allocated for GA100 or switch to Samsung's 8LPP instead.

There aren't enough N7 wafers to go around, getting ahold of enough supply for all their consumer GPUs is difficult. That being said I wouldn't be surprised if they use N7 for at least one of their consumer dies, even if it's a retape they spin after the initial lineup rolls out.

I suspect this is also why the rumour that Nvidia are buying N5 wafers early has come about. They don't want to be in this situation again.

I can say seperately I'm confident that Samsung ran into major issues with 7LPP, because I know that it had a knock on effect with their 5nm later down the line. The issue was likely tackled afaik... but you don't just fix an issue like this without knock on effects. And there were definitely knock on effects.

Okay, so what about the official statement that TSMC would be producing the bulk of NV GPUs?

 

[Krteq]

Yes, NV will manufacture some GPUs at TSMC, but they also had to manufacture some GPUs at Samsung because there is no 7nm capacity left at TSMC.

 

[uzzi38]

Okay, so what about the official statement that TSMC would be producing the bulk of NV GPUs?

No clue, I'm mostly just guessing with the info I've seen or getting. That comment is the only reason I've still even slightly skeptical of Kimi, and have been entertaining the thought of GA102 or GA104 on N7.

At this point, who knows?

 

[DisEnchantment]

Yes, NV will manufacture some GPUs at TSMC, but they also had to manufacture some GPUs at Samsung because there is no 7nm capacity left at TSMC.

Who has picked up HiSilicon's allocation of almost 30K wpm at TSMC?
On top of that, with the smartphone business going down there are lots of wafers available

TSMC is now concerned with unused capacity 😕

 

[uzzi38]

Who has picked up HiSilicon's allocation of almost 30K wpm at TSMC?
On top of that, with the smartphone business going down there are lots of wafers available

TSMC is now concerned with unused capacity 😕

That's an interesting point with the developments over the last couple of days.

I don't think that supply will have gone immedietly though. Lets wait and see how the next few days go, but to be fair - it's not like Nvidia could have anticipated that there would be 30k wpm that would suddenly become free in the space of a week. Especially not ~6 months before they were planning on releasing their GPU lineup.

 

[jpiniero]

That's an interesting point with the developments over the last couple of days.

Intel is who came to mind.

SS 8 btw is a decent shrink compared to 12FFN. It's just not comparable to either foundry's 7 nm node.

 

[sontin]

NV tried to manipulate TSMC for better prices, so they taped out some chips at Samsung... but their risk didn't ended well... other companies took their opportunities and split up a rest of 7nm capacity from leaving Apple. So NV had to manufacture some chips at Samsung because there are no capacity left at TSMC

They are producing 827mm^2 dies. Money is not even remote one of their worries. And everyone is paying the same price.

 

[Konan]

So, unless I am freaking blind, I didn't see any mention of Samsung being the target fab. (Edit, should have clicked on second link - still, what in the freaking hell is NV doing?!).
At this point in time, it would only make sense to target 7nm at either SS or TSMC.

Also, I think tapeout to retail availability is typically around a year, give or take. I suppose NV anticipated this and has added resources to reduced the TTM. Anyway, really bad news for NV if true.

https://twitter.com/x/status/1237687606869188608

It's important to note with Kopite7kimi twitter original post was MAY 2019 not this year. His second add to it was March 2020. So approximately 1 year ago started the all the Ampere gaming RTX cards will be coming from Samsung - a lot can change since then...

 

[Ajay]

It's important to note with Kopite7kimi twitter original post was MAY 2019 not this year. His second add to it was March 2020. So approximately 1 year ago started the all the Ampere gaming RTX cards will be coming from Samsung - a lot can change since then...

Yeah, and reading that post again he says "It was based on Samsung 10nm". So he implies a change but doesn't specifically say so. I guess he must be unsure. So, we are stuck waiting for confirmation on which particular GPU will be manufactured where. Bummer.

 

[Stuka87]

And here some people said a few of us were crazy for saying all the gaming cards were going to be at Samsung, and only HPC was going to be at TSMC.

 

[Ajay]

And here some people said a few of us were crazy for saying all the gaming cards were going to be at Samsung, and only HPC was going to be at TSMC.

I think Nvidia was using it's marketing dominance to play a dangerous game. Now that they realize AMD is actually betting heavily on their GPU lineup, NV woke up and booked a bunch of 5N EUV wafers so they don't get behind the eight ball. Interesting times.

 

[Konan]

And here some people said a few of us were crazy for saying all the gaming cards were going to be at Samsung, and only HPC was going to be at TSMC.

It's not crazy, 14 months ago leaks came out saying Samsung 7nm (someone else) and right up till today only that Kopite guy has said the full line gaming lineup is on Samsung. I still don't think the entire gaming stack will be on Samsung but even if it is I am sure it will still be competitive. It is not like it is an issue at present. Nvidia has for several years had superior execution compared to AMD gpus. No reason to for me to think that is going to change any time soon, but competition is welcomed. I'm in the market, like a suspect a few people are and will wait for those in depth reviews before making a purchase choice.

 

[DXDiag]

Who has picked up HiSilicon's allocation of almost 30K wpm at TSMC?
On top of that, with the smartphone business going down there are lots of wafers available

TSMC is now concerned with unused capacity

Exactly, there is now more 7nm capacity for everyone, we've already concluded that several pages ago.

 

[darkswordsman17]

It's important to note with Kopite7kimi twitter original post was MAY 2019 not this year. His second add to it was March 2020. So approximately 1 year ago started the all the Ampere gaming RTX cards will be coming from Samsung - a lot can change since then...

I'm not going to say its impossible, but I don't think they could/would have made that change in even a year's time. It would require pretty extensive changes to the design of the chip (I'm not sure it would require a complete change but its not something that would be easy, and they'd need to know well in advance of finalizing the engineering of the chip what process they were going for).

The other problem with that is that Nvidia wouldn't have known back then that the smartphone market would expect slowdowns and cut wafer buys. They'd have had only what 2 months to make a decision based on that information.

Which I'm not saying Nvidia won't have 7nm gaming chips. I really don't know. But if they already had them planned for Samsung's 10/8nm, and even allegedly having had them taped out or in the process of it, then its very unlikely that'd change just because TSMC has extra capacity. Nvidia could probably use that for their HPC stuff.

 

[Veradun]

Interesting. Anyone hazard a guess if GA100 will show up in RTX/Titan cards or is that for Tesla/Quadro only?
I'd guess GA102 -> RTX 3080 Ti but maybe they'll bring GA100 as a Titan card later. It's hard to guess because they did release both GV100 and TU102 as Titan cards.

Nope. They will cut down A102 to enable the meme tier Titan A on the full chip. If yelds are a concern then the Titan A might also be a cut down version (just less cut), enabling the super meme tier Titan Ax. Depending on AMD's offer they can also choose to only market the Titan, leaving the Ti for later, for more milking.

No clue, I'm mostly just guessing with the info I've seen or getting. That comment is the only reason I've still even slightly skeptical of Kimi, and have been entertaining the thought of GA102 or GA104 on N7.

At this point, who knows?

A102 is definitely possible, A104 seems to have too much volume for their seemingly available capacity at TSMC.

They are producing 827mm^2 dies. Money is not even remote one of their worries. And everyone is paying the same price.

Those dice are selling for a crapload of money, I wouldn't infer their will to lose money from that.

 

[sontin]

Those dice are selling for a crapload of money, I wouldn't infer their will to lose money from that.

To lose money on <400 dies and a price tag of a card of $400 would mean that nobody could produce any consumer GPU on TSMC's 7nm process...