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.

 

[sontin]

nVidia knows everything about every process. Samsungs 8nm process is in mass production since summer 2018. Gaming-Ampere had his tapeout six months ago or so.

 

[Gideon]

EUV is a different beast. A100 is still 7nm DUV. So it makes even less sense to go with Samsungs 7nm process.

Samsung only had a EUV based 7nm process, and it initially and it was supposed to be ready ages ago (Risk production Oct 2018 First own SOC: Exynos 9825 in August 2019, supposed mass production for all in Feb 2020).

Yet it sees it failed to meet Nvidia's needs (large high-TDP chips) So they offered the 10LPE based 8nm process instead:

Samsung offered nVidia a pretty good deal... had they delivered.

Exactly

nVidia knows everything about every process. Samsungs 8nm process is in mass production since summer 2018. Gaming-Ampere had his tapeout six months ago or so.

What are you hinting here exactly? Obviously 8nm was chosen at least a year ago, as Samsung 7nm failed to deliver and by that time TSMC was booked.

Had Samsung's EUV 7nm worked as intended, it would have been an excellent process. It did not.

 

[Glo.]

What are you hinting here exactly? Obviously 8nm was chosen at least a year ago, as Samsung 7nm failed to deliver and by that time TSMC was booked.

He is suggesting that using 8 nm Samsung process was very thouroughly thought business decision on the Nvidia side.

 

[beginner99]

It just doesn't feel right for them to release such a huge power sucking card without another motive.

Isn't the motive obvious? To not tarnish the "Titan" brand name as a hot, big mess.

 

[Gideon]

He is suggesting that using 8 nm Samsung process was very thouroughly thought business decision on the Nvidia side.

Good point. Its obvious the decision had to come a while ago (probably at least 1,5 years) and by that time that node's perfomance was well known. That means Nvidia was confident they would be able to offer enough of an improvement over Turing and whatever RDNA2 brings (they obviously couldn't have known much about the latter though).

 

[Glo.]

Good point. Its obvious the decision had to come a while ago (probably at least 1,5 years) and by that time that node's perfomance was well known. That means Nvidia was confident they would be able to offer enough of an improvement over Turing and whatever RDNA2 brings (they obviously couldn't have known much about the latter though).

No.

Nvidia wanted to use Samsung's 7 nm EUV for ALL of their products. From A100, 102, 103, 104, 106, 107 dies.

ALL OF THEM.

Now we end up in situation in which A100 is on 7 nm TSMC process, and gaming cards are on 8 nm SS process.

Why? Because of huge business decision blunder.

 

[Gideon]

No.

Nvidia wanted to use Samsung's 7 nm EUV for ALL of their products. From A100, 102, 103, 104, 106, 107 dies.

ALL OF THEM.

Now we end up in situation in which A100 is on 7 nm TSMC process, and gaming cards are on 8 nm SS process.

Why? Because of huge business decision blunder.

Why "No"? My point was that this decision couldn't have come much later than ~1 year ago (simply because of the time it takes to redesign to other nodes) probably a bit earlier. I never disputed their initial plans (for Samsung EUV). Nor did I ever say that decision was a good one.

 

[sontin]

What are you hinting here exactly? Obviously 8nm was chosen at least a year ago, as Samsung 7nm failed to deliver and by that time TSMC was booked.

Had Samsung's EUV 7nm worked as intended, it would have been an excellent process. It did not.

TSMC has a 7nm EUV process, too. And nobody is using it. Apple and co. are jumping straight to 5nm EUV. EUV was not relevant for nVidia. And Samsungs 7nm EUV process isnt even so much better than the 8nm process.

 

[Saylick]

And Samsungs 7nm EUV process isnt even so much better than the 8nm process.

I wouldn't be so sure about that. Samsung's 8LPP is still based off of their 10nm process but with a density increase, while their 7nm and its derivatives are full-fledged EUV nodes. Even if we ignore the benefits of the density improvement that EUV provides, we can expect a non-insignificant transistor performance improvement between the 10nm family and 7nm family of nodes, likely >15% better perf @ iso-power if we go by TSMC's numbers (i.e. N7 to N5).

EDIT: Corrected N7+ to N5, seeing as how that would be the equivalent jump from TSMC's N7 DUV-only node to their full fledged EUV node.

Also, if we do consider the density improvement that Samsung's 7LPP would have provided over their 8LPP, even if Samsung was offering Nvidia 1/3rd cheaper wafers, that reduction in density would have undone most of that price difference as there's almost a 50% difference in max density between those two nodes. If we acknowledge that logic scales less than cache, there is still probably a 15-20% inflation in die sizes that Nvidia had to eat as a result of using 8LPP.

 

[Ajay]

No.

Nvidia wanted to use Samsung's 7 nm EUV for ALL of their products. From A100, 102, 103, 104, 106, 107 dies.

ALL OF THEM.

Now we end up in situation in which A100 is on 7 nm TSMC process, and gaming cards are on 8 nm SS process.

Why? Because of huge business decision blunder.

I’d say Samsung is as much at fault as Nvidia, since they over promised and under delivered.

 

[Shivansps]

This feels like as Polaris launch, with RX480 clocked way too high, when the 1060 launched we realised why they had to do that.

 

[tajoh111]

No.

Nvidia wanted to use Samsung's 7 nm EUV for ALL of their products. From A100, 102, 103, 104, 106, 107 dies.

ALL OF THEM.

Now we end up in situation in which A100 is on 7 nm TSMC process, and gaming cards are on 8 nm SS process.

Why? Because of huge business decision blunder.

It is way to early to call it a blunder without performance and actual sales being released.

One aspect you need to take into account is supply and volume.

Nvidia's cost of good sold is the same as AMD's which means they spend a similar amount of wafers.

Adding Nvidia demand on wafers means TSMC would not be able to provide them the necessary wafers to provide the volume of product Nvidia needs which is enormous.

As seen with recent marketshare figures, the public is still willing to buy Nvidia even when AMD has a lower price and is competitive with navi as shown this quarter with Nvidia having an 80% marketshare in the discrete segment.

https://wccftech.com/nvidia-geforce-gpus-gain-major-market-share-versus-amd-radeon-in-q2-2020/

So even if AMD is more competitive this time around, when you combing the savings on wafers, the volume of wafers Samsung can provide, it might be the more optimal business decision when you consider the volume and price flexibility going with samsung will likely give them. E.g, Having 4x the supply of product along with a lower cost for nvidia which allows them to sell for cheaper, use larger dies or absorb more profit from each sale.

It is too early to call this a blunder Akin to Vega or Radeon VII which both had to be sold initially at cost, along with their poor volume made them poor profitability.

Business decisions are measured on financial success not personal biases(e.g jumping on negative rumors for a company you don't like and magnifying positive rumors[ignoring negative ones] for the company you like).

 

[sontin]

I wouldn't be so sure about that. Samsung's 8LPP is still based off of their 10nm process but with a density increase, while their 7nm and its derivatives are full-fledged EUV nodes. Even if we ignore the benefits of the density improvement that EUV provides, we can expect a non-insignificant transistor performance improvement between the 10nm family and 7nm family of nodes, likely >15% better perf @ iso-power if we go by TSMC's numbers (i.e. N7 to N5).

7nm EUV was ~15% better than 8nm LPP: https://www.anandtech.com/show/1560...xynos-snapdragon-review-megalomania-devices/9
Samsung failed to deliver a competitive 7nm process.

 

[uzzi38]

7nm EUV was ~15% better than 8nm LPP: https://www.anandtech.com/show/1560...xynos-snapdragon-review-megalomania-devices/9
Samsung failed to deliver a competitive 7nm process.

Exynos 990 is 7LPP, not 8LPP.

 

[sontin]

And Exynos 9820 is on 8nm LPP. Neither the CPU nor the GPU has made a huge perf/watt jump.

 

[uzzi38]

And Exynos 9820 is on 8nm LPP. Neither the CPU nor the GPU has made a huge perf/watt jump.

That's correct. Wasn't really disagreeing, though looks like I misread the context, so excuse my poor add in.

It's true that 7LPP vastly underperformed compared to it's targets however.

 

[DDH]

My understanding is that DUV and EUV are not ip compatible, and require at least some redesign. That A100 is on TSMC 7nm DUV to me at least indicates that the ampere line was always designed for DUV and not EUV. I believe NVIDIA changed messages regarding its ampere manufacturing node at least once, from Samsung being the main fab, to TSMC supplying all 7nm products. From this I would guess that if Samsung 8nm was not proving as good as first projected, NVIDIA shifted what they could to TSMC, the HPC cards being priority. And with limited TSMC capacity due to AMD hoovering up a lot, that chip is all they could get manufactured there.

Sent from my SM-N975F using Tapatalk

 

[Stuka87]

Adding Nvidia demand on wafers means TSMC would not be able to provide them the necessary wafers to provide the volume of product Nvidia needs which is enormous.

nVidia is a tier 3 (soon to be 4) TSMC customer. The number of wafers they use pales in comparison to Tier 1-2 customers. Because nVidia is a lower tier, they have less priority, and possibly pay more. And like I mentioned in another post, nVidia tried to play hard ball with TSMC and lost. So now they are stuck with Samsung 8nm.

 

[MrTeal]

nVidia is a tier 3 (soon to be 4) TSMC customer. The number of wafers they use pales in comparison to Tier 1-2 customers. Because nVidia is a lower tier, they have less priority, and possibly pay more. And like I mentioned in another post, nVidia tried to play hard ball with TSMC and lost. So now they are stuck with Samsung 8nm.

Is there a source on that or somewhere that lists TSMC's customers by tier?

 

[maddie]

nVidia is a tier 3 (soon to be 4) TSMC customer. The number of wafers they use pales in comparison to Tier 1-2 customers. Because nVidia is a lower tier, they have less priority, and possibly pay more. And like I mentioned in another post, nVidia tried to play hard ball with TSMC and lost. So now they are stuck with Samsung 8nm.

That was truly a strategic blunder. I'm sure that TSMC will long remember this and be very wary of allowing them any sort of unnecessary leverage soon. Cooperative partnerships is what allowed TSMC to be what they are today and it seems that Nvidia loves antagonistic ones.

 

[kurosaki]

afers they use pales in comparison to Tier 1-2 customers. Because nVidia is a lower tier, they have less priority, and possibly pay more. And like I mentioned in another post, nVidia tried to play hard ball with TSMC and lost. So now they are stuck with Samsung 8nm.

Nvidia, soon fabbed on a GloFo fab near you! (TM)

 

[Saylick]

That was truly a strategic blunder. I'm sure that TSMC will long remember this and be very wary of allowing them any sort of unnecessary leverage soon. Cooperative partnerships is what allowed TSMC to be what they are today and it seems that Nvidia loves antagonistic ones.

Yeah, that is something I always never liked about Nvidia and I am honestly glad they got their hand slapped, especially when it's the one that feeds them. JHH must have been salivating at the thought about improving margins by using Samsung but Nvidia literally couldn't have picked a worse company to bully. I wonder if TSMC could legally refuse to do business with Nvidia entirely... I'm sure that ought to make JHH to beg and grovel to get back on TSMC's good graces.

 

[n0x1ous]

Yeah, that is something I always never liked about Nvidia and I am honestly glad they got their hand slapped, especially when it's the one that feeds them. JHH must have been salivating at the thought about improving margins by using Samsung but Nvidia literally couldn't have picked a worse company to bully. I wonder if TSMC could legally refuse to do business with Nvidia entirely... I'm sure that ought to make JHH to beg and grovel to get back on TSMC's good graces.

Well that obviously didnt happen since TSMC is making A100 for them on 7nm right now

 

[Ajay]

Yeah, that is something I always never liked about Nvidia and I am honestly glad they got their hand slapped, especially when it's the one that feeds them. JHH must have been salivating at the thought about improving margins by using Samsung but Nvidia literally couldn't have picked a worse company to bully. I wonder if TSMC could legally refuse to do business with Nvidia entirely... I'm sure that ought to make JHH to beg and grovel to get back on TSMC's good graces.

From what I read, there was no bullying involved (at least initially). Samsung offered a good deal from the get go.

 

[blckgrffn]

Nvidia, soon fabbed on a GloFo fab near you! (TM)

Strategic alliance between IBM and Nvidia to move GloFo forward? Hahahaha, that'd be rich.

Speaking of which...

TSMC Launches New N12e Process: FinFET at 0.4V for IoT

www.anandtech.com

Appears that there is a lot of 12nm fab capacity open at GloFo?