NVIDIA Allegedly Moving Ampere to 7nm TSMC in 2021

Lakados

2[H]4U
Joined
Feb 3, 2014
Messages
2,717
Between the retooling costs and TSMC’s existing contracts I find this hard to believe. Perhaps as a late 2021 refresh for a limited batch of Super/TI cards for announcement in October hard maybe but outside of that I just don’t see it. By the time they get there Samsung should have their house in order and yields shouldn’t be an issue any more.
 

Ready4Dis

2[H]4U
Joined
Nov 4, 2015
Messages
2,332
AMD rumour mills.
I don't think it's coming from AMD rumor mills... I'm pretty sure the lack of cards and the fact that even NVIDIA admitted they won't have real stock to meet demand until 2021 played into that a bit ;). You don't run out of stock because you have so many chips available and nothing to do with them. There were rumors/concerns PRIOR to launch about Samsung yields as well, so it's not as if some amd fan boy came out of the blue to say this. I don't know what samsung yields are and we may never know, but one things for sure, NVIDIA is not making as many cards as they were hoping to and all the marketting/spin isn't going to change that. Whether they jump ship next year or not will be a pretty good indication on whether they were happy with the yields vs. price or not.
 

DanNeely

2[H]4U
Joined
Aug 26, 2005
Messages
3,700
I find this rumor hard to believe. Retooling it to work with TSMC wouldn’t be cheap unless NVIDIA feels Navi 21 is too big of a threat so they’re creating 7nm super cards on tsmc. Either way I don’t see that happening till at least June if not longer.
NVidia's been splitting production between Samsung and TSMC for a while now; I assume all their designs are being validated on both companies processes and then at the end they decide who to use based on price/performance/availability numbers.
 

harmattan

Supreme [H]ardness
Joined
Feb 11, 2008
Messages
4,564
Between the retooling costs and TSMC’s existing contracts I find this hard to believe. Perhaps as a late 2021 refresh for a limited batch of Super/TI cards for announcement in October hard maybe but outside of that I just don’t see it. By the time they get there Samsung should have their house in order and yields shouldn’t be an issue any more.
Retooling and negotiating contracts for a mid-cycle refresh that will likely stomp all over AMD (when they'll likely still be designing RDNA 3), and allow them to get their feet wet on 7nm on a proven architecture is advantageous in many ways.

Well wow, ya don't say...

https://hardforum.com/threads/rtx-3xxx-performance-speculation.1987972/page-86#post-1044726608
 
Joined
Apr 22, 2011
Messages
514
Glad the nvidia debut was a trainwreck, now everyone knows to not only wait for AMD but also wait for the refresh for a better 3000 series gpu that doesn't draw as much power and better performance.
 

exlink

Supreme [H]ardness
Joined
Dec 16, 2006
Messages
5,041
I find the rumor to be plausible but unlikely. As already mentioned by others, retooling is costly and takes a long time. Samsung yields would need to be absolute garbage to consider making the switch, even if it’s for a “Super” refresh.

The only reason I find it plausible is that Apple moved to 5nm on its newest phones which likely opened up a lot of room for 7nm. TSMC may try to get Nvidia back for Ampere instead of waiting for Hopper by offering them pricing lower than what was previously negotiated.
 

DeChache

Supreme [H]ardness
Joined
Oct 30, 2005
Messages
7,076
I find this rumor hard to believe. Retooling it to work with TSMC wouldn’t be cheap unless NVIDIA feels Navi 21 is too big of a threat so they’re creating 7nm super cards on tsmc. Either way I don’t see that happening till at least June if not longer.


This ^^^^^ you just dont move a chip to a new fab and process. The chips are designed with both fab capability and process in mind not to mention the NDAs where usually a team can only have access to one fabs design docs.

A super refresh next summer on TSMC sure. Moving exciting cards that would mean they designed for both fabs from the get go.

Now the A100 is already TSMC so maybe enough of the design is done they can do it quickly but quickly to me is still spring.
 

defaultluser

[H]F Junkie
Joined
Jan 14, 2006
Messages
13,720
The sever lack of product does.


Have you forgotten the three months of non-availability for the tiny die size of the 314 mm² GTX 1080 back on TSMC 16nm? Even though they price it initially at $700 to lower the expected demand?

This is the first mass-produced large card on this fine a process node, so of course yields are shit. It's going to take 3 months after retail release (Janary) before stock should return.

You also have other shortages of everything else you need to male these cards...for example, GDDR6x chips that are brand-new, and don't grow on trees,
 
Last edited:

Aireoth

Supreme [H]ardness
Joined
Oct 12, 2005
Messages
4,242
Have you forgotten the three months of non-availability for the tiny die size of the 314 mm² GTX 1080 back on TSMC 16nm? Even though they price it initially at $700 to lower the expected demand?

This is the first mass-produced large card on this fine a process node, so of course yields are shit. It's going to take 3 months after retail release (Janary) before stock should return.

You also have other shortages of everything else you need to male these cards...for example, GDDR6x chips that are brand-new, and don't grow on trees,

I have forgotten nothing, but I am also aware that actual ampere cards shipped to NA whete around 10k at launch, a historic low. Keep that head in the sand, or shill, whatever floats your boat.
 

defaultluser

[H]F Junkie
Joined
Jan 14, 2006
Messages
13,720
I have forgotten nothing, but I am also aware that actual ampere cards shipped to NA whete around 10k at launch, a historic low. Keep that head in the sand, or shill, whatever floats your boat.

Well, then you're just ignoring reality, and whining senselessly like a 3 year old. Complete fucking waste of my time, I'm out..
 
Last edited:

defaultluser

[H]F Junkie
Joined
Jan 14, 2006
Messages
13,720
If Apple moved to 5nm, then why wouldn't Nvidia just move to 5nm?

EDIT: And further, I think it is clear that Ampere doesn't have the headroom for a refresh and that Nvidia needs a major architecture change to even get a 10% performance boost.


The reason why Nvidia has stopped mid-architecture die shrinks is because they no-longer provide the substantial clock seed boosts/die size reduction they used to. When you can typically get a 10%-boost at the same power level on a refined process, what's the point of a 20% die shrink? And also, they can cheat with reducing the prices of larger core cuts.

It's also become ungodly expensive to go back to doing a die shrink every two years. Just ask Apple how fucking much they have to charge for cell phones, and Apple Store 30% cut to be able to afford all that?

Nvidia doesn't sell hundreds of millions of GPUs every year, but Apple does. But they are getting their start on TSMC 5nm for their next Tegra chip (6 months after Apple has worked all the bugs out of the process). It's going to be another year before we hear anything as large as the RTX 3090 being produce on 5nm. Even Apples Mac ARM processors are going to be 12 total cores (probably two to three times the die size of their tiny cell phone chip.)
 
Last edited:

Teenyman45

2[H]4U
Joined
Nov 29, 2010
Messages
2,523
To those are insisting on the amount of retooling involved, Nvidia has already been producing at least one Ampere chip on TSMC's 7nm process for some time. Though, I do agree this would like be for a refresh of Ampere "supers" to benefit from the power savings to be expected on TSMC's 7nm process. Just look at how EVGA is releasing a 450w beta BIOS for one of the 3080 FTW cards.

I still prefer the argument that Nvidia went to Samsung not because TSMC absolutely lacked the wafer capacity for Nvidia, but because Nvidia was trying to get a lower wafer price and TSMC wouldn't knuckle under... this has then led to poorer than expected results for Nvidia, especially if Big Navi (or supposedly a higher CU count / faster GPU that I nickname "Biggest" Navi) is finally outright competitive with Nvidia in the higher end space for the first time since Nivida took the undisputed lead with the 780ti and Titan.

One thing is clear, Samsung can't produce the Ampere-based GPUs quick enough or in enough volume so, from the looks of things, NVIDIA will be moving a large chunk of wafer fabrication to TSMC...

https://www.techpowerup.com/273302/nvidia-reportedly-moving-ampere-to-7-nm-tsmc-in-2021

waiting to buy a 3080 sounds like it might turn out to be the best move...

Then on the other side of can't produce enough GPUs quick enough are people like Mr. Moore's Law is Dead who has been alleging that Nvidia had been stockpiling GPUs and especially GDDR6x for a while so that there would be a big flood of mostly elevated price 3080's during November or December to profit off of the artificial scarcity and try to prevent AMD from gaining market share with Big Navi.
 

defaultluser

[H]F Junkie
Joined
Jan 14, 2006
Messages
13,720
Then on the other side of can't produce enough GPUs quick enough are people like Mr. Moore's Law is Dead who has been alleging that Nvidia had been stockpiling GPUs and especially GDDR6x for a while so that there would be a big flood of mostly elevated price 3080's during November or December to profit off of the artificial scarcity and try to prevent AMD from gaining market share with Big Navi.

Do you even understand how stupid this accusation sounds?

How is a an overpriced RTX 3080 somehow going to prevent Big Navi from destroying NVIDIA with their superior numbers and prices.

Assuming you don't believe the hype from AMD, and don't believe they will have massive qualities available?
 

Ready4Dis

2[H]4U
Joined
Nov 4, 2015
Messages
2,332
Then on the other side of can't produce enough GPUs quick enough are people like Mr. Moore's Law is Dead who has been alleging that Nvidia had been stockpiling GPUs and especially GDDR6x for a while so that there would be a big flood of mostly elevated price 3080's during November or December to profit off of the artificial scarcity and try to prevent AMD from gaining market share with Big Navi.
Did he say stockpiling ever? I only recall (admittedly I haven't looked back at it) that he said there wouldn't be much availability @ MSRP and most cards @ MSRP would be inferior (do capacitor issues count?) to FE (which will be barely available) and other models will cost more. I don't recall him mentioning stockpiling or anything but it's been a while so I may have missed/forgot about it by now. It was mostly speculation (as are most of these youtube starts), but that doesn't mean he's completely wrong either (or 100% right), if you throw enough stuff at a wall, some of it is bound to stick ;). Either way, he seems more right than wrong so far, especially with NVIDIA pawning off FE cards to Best Buy and telling the EU they can't even buy them anymore and we haven't seen to many MSRP cards and the ones that were have been lower quality and needed a boost clock fix in order to stop crashing (make it slower ramp up/down due to possible capacitor/filter issues). Either way, I never expect any rumor to be 100% accurate unless it's something like a leak of an official document or spec, but like I said, so far it seems much more right than wrong, and now with the rumors of 20gb cards coming before we even have availability of regular cards, it seems like this may have always been the direction. Release cheap FE card, get good benches @ "MSRP" and then we all end up with cards over MSRP (either 10gb or 20gb).
 

Axman

Supreme [H]ardness
Joined
Jul 13, 2005
Messages
5,996
I don't remember anyone with MLID saying Nvidia was stockpiling glass for this. If anything the main theme recently has been Guys, Wait for Big Navi.

I think things point to Nvidia developing enterprise hardware with TSMC, which would still be a big deal, if not for gamers, and I expect them to move forward with Samsung for consumer graphics. It's clear that Ampere was pushed outside of the "sweet spot," but that doesn't rule out better performance per watt in the future, especially on better nodes.

It takes, what, a year to tool up? If this rumor got out six months ago, then maybe it would seem more reasonable. But for now I think Nvidia is going to make do with and continue to work with Samsung; clearly they have something going right, even if they pushed things -- both in terms of performance and in terms of yields -- for this launch.
 

Axman

Supreme [H]ardness
Joined
Jul 13, 2005
Messages
5,996
Oh I thought we were talking about the just GPU production. Still, they can stockpile all the RAM they want; they have to get the main die yields up. For their sake and Samsung's.
 

Kardonxt

2[H]4U
Joined
Apr 13, 2009
Messages
3,238
images.jpg
 

Sycraft

Supreme [H]ardness
Joined
Nov 9, 2006
Messages
4,722
So... What evidence is there that TSMC 7nm is so much better than Samsung 8nm? I feel like people see the smaller number and go "Oh it must be more advanced!" Remember that these days what they tag the process size with is marketing, it has little to do with reality. Different companies name their shit differently. So what actual evidence is there that TSMC 7nm is better? I'm not saying it isn't but let's see some factual comparison not just "But 7 is smaller than 8!"
 

ChadD

Supreme [H]ardness
Joined
Feb 8, 2016
Messages
4,772
I can see them moving to TMSC for a super variant. If Nvidia is willing to pay the premium. TMSC has no need to make anyone deals. For anyone saying Nvidia wouldn't pay the price to retool ect... of course they would. They are sweating bullets right now. They really really don't want to have to take the Ego hit being number 2 for even a few months. Even if AMDs big navi isn't faster... and I tend to believe it will be close but not quite more. I get the feeling they are very worried AMD would be close enough to release some binned super type chip of their own to at least grab the crown at the insanity price point.

I'm also not shocked if Samsung yields are as bad as the rumors... and the rumors seem to be pretty well borne out by the actual stock, IF Nvidia didn't go months early for a marketing win. I still believe Nvidia probably just "launched" 2-3 months too early as soon as they got working chips... where as normally they would have built stock for at least a few months to have well supplied launch. Samsung lets all remember has never fabbed anything like Ampere before. There is a HUGE difference between (pun intended) between a 127nm exynos chip and a 826nm Nvidia GPU. The reports have always been that Nvidia had issues with yields due to its massive die sizes.... its hardly shocking that Samsungs yields would be iffy as they have never had experience with anything even mid range size never mind massive size with 1/4 the number of chips per wafer.
 

Ready4Dis

2[H]4U
Joined
Nov 4, 2015
Messages
2,332
They are all marketing numbers by now, 8nm != 8nm and TSMC 7nm != 7nm.

It's hard to directly compare processes, but I'll give it a try...

Samsung 8nm LPU uHD (LPP is for low frequency, LPU is for high frequency, uHD is ultra high density, HD is normal high density)
Fin Pitch - 42nm
Gate Pitch - 64nm
M1, Mx - 44nm
Density - 61.18MTr/mm2

TSMC 7nm HP (high performance, LP is for low power and actually has better density, but slower)
Fin Pitch - 30nm
Gate Pitch - 64nm (For low power, this is only 57nm, hence ever more dense than HP, but it can't hit as high of frequencies)
M1, Mx - 40nm
Density - 91.2MTr/mm2


So, as you can see, TSMC has higher densities by a pretty good margin. This also doesn't completely take into account efficiencies of the exact materials/process either, but gives you a good indication why TSMC 7nm is generally considered better (it's almost 50% more dense!). There are also all sorts of different designs and such, I tried to find the ones that best represented the current GPU processes but there may be slight differences.
 

toast0

Gawd
Joined
Jan 26, 2010
Messages
1,009
It was no secret that Nvidia went Samsung because AMD bought out the TSMC capacity. So capacity would have to be freed up in order for this rumor to be true. So what is being freed up?
I was seeing speculation that Huawei was getting as much fab time as possible before the latest round of restrictions went into force. Sounds plausible, but I can't find a good source for the date the restrictions took effect.
 
Top