Overclock.net banner

[GURU3D] NVIDIA Allegedly Moving Ampere to 7nm TSMC in 2021

9K views 42 replies 23 participants last post by  doom26464 
#1 ·
-- Digitimes (online translated) --
It has also been reported that NVIDIA’s previous annual masterpiece RTX 30 series uses Samsung Electronics’ 8nm process at a considerable OEM discount, but will switch to TSMC’s 7nm process in 2021. The order volume is not less. The reason for NVIDIA's transfer of orders is that TSMC's 7nm offer is relatively close to the people, and the other is that it has previously planned to diversify risks to deal with Samsung's 8nm yield problem. The NVIDIA 7nm large order is also one of TSMC's important customers to maintain high-end 7nm capacity utilization in 2021.
source
 
  • Rep+
Reactions: vinni3pags
#3 · (Edited)
Could kind of see this coming, considering Big Navi will be near or matching RTX 3080 performance with significantly smaller die and power consumption. I was trying to explain to people how good RDNA 1 was for its performance to die size ratio.... and that they have yet to release a high end GPU when looking at their die size.

If you look at some of my old post i was doing the math of where RDNA 2 was going to be at just by doubling the cores and increasing the die size.

The 5700XT has a die size of 251mm and look at its performance. When comparing it to the performance of the RTX 2080ti @ a 775mm die. RDNA 1 was clearly way head of its time as far as die to performance ratio. Now people are surprised at the performance coming from RDNA 2 on a 500+mm die... If AMD took the lets make a big die route. They would of been ahead of Nvidia for close to 2 generations when RDNA was first released.

Over all i think big Navi will be an over all better card. Even if it is a little slower, it will be running cooler, using less power and almost matching their competitors performance with a smaller die. Also they have been working on it for quite a bit longer than Nvidia and their RTX 3000 series. Also RDNA will likely be able to overclock to around 2.6 Ghz

With EPYC, RDNA 2, Xbox, PS5 and Zen3 iam predicting a massive shift in revenue to AMD. Now would be a good time to invest in AMD stocks, they will probably start to soar very soon. If rocket lake isnt able to steal back the single threaded crown, AMD will likely be the new empire.
 
#7 ·
Now all they have to do is finally beat the 2070 super. At whatever the die size.
 
#4 ·
That rumor has been travelling around for couple days. It was something that moore's law is dead mentioned months ago. I have my doubts though. If true, people will wait for TSMC 7nm, at least for the power savings. I think the one's who struck out didn't really lose since this was real launch to avoid.
 
#5 ·
I dont bite on this rumour and can see nvidia using samsung for the 2 year life cycle of the 3000 series.

We wont get a change till 4000 series which nvidia may shift back to tsmc or something else(maybe more samsung?)

I cant see them switch to a whole different node and foundry mid product cycle for an already taped out arch. I mean anything is possible but with such high demand anyways I dont think nvidia has really any push to make a better product at this point.
 
#9 ·
I dont bite on this rumour and can see nvidia using samsung for the 2 year life cycle of the 3000 series.

We wont get a change till 4000 series which nvidia may shift back to tsmc or something else(maybe more samsung?)

I cant see them switch to a whole different node and foundry mid product cycle for an already taped out arch. I mean anything is possible but with such high demand anyways I dont think nvidia has really any push to make a better product at this point.
It sounds like the move, if it happened, would be to address concerns with price and supply shortage. You're right though, they can't exactly copy paste to a different node.
 
#6 ·
Is this the same Digitimes that said AMD would delay zen 3 to 5nm because they didnt have to rush? i know this is a rumor but dear god at face value its junk. You cant take a Samsung 8Nm chip and put it on a TSMC 7Nm without a lot of re engineering despite what Digitimes keeps trying to convey. I personally dont see AMD beating the 3080 but i do see it getting within spitting distance and if so then Nvidia would adjust pricing/change memory configurations.
 
#8 ·
Moving to a new node is stupid and will affect performance, power consumption, temperatures, etc. It's not happening.
 
#10 ·
Yeah its odd. I would imagine Nvidia would shrink PCB too if they had to move to TSMC 7nm. Lot of pissed off buyers.
 
#15 ·
They probaly are indeed buying capacity in 2021 for TSMC for 7nm.

But its probaly for there 4000 series that will launch in 2022 aka hopper.

Foolish to assume that its for a mid product refresh on an entire different node. Judging by digitimes track record though what do you except.
Hopper tsmc 5nm i thought.
 
  • Rep+
Reactions: SmartBombs
#14 ·
I doubt they can port the whole line mid-generation, but given how poorly the launch has gone and how there's no supply anyway, what's really the difference?
 
#16 ·
I doubt nVidia would do this mid-generation unless their production came to halt due to manufacturing problems. If they did that, series 3000 and nVidia would get a far more damaging reputation than simply a "paper launch".
 
#18 ·
Consumer stuff has always ment to be 8nm samsung and the server line up later on 7nm tsmc since ampere was announced.
There is a fab agreement that your not allowed to use the same chips between different fabs so you wont be seeing 3080 and 3090s on 7nm.

But what you might see is 3080 and 3090 supers which has the memory clocked higher once micron has gddr6x more mature and in 2GB chips.
Then a 3090ti or titan thats tsmc 7nm from the cut down server stuff which is like 280W and beats the 3090 at 480W.
 
#20 ·
Good news! if Nvidia can do a fresh with higher clocks and more ram and issue something that is about 60-70% above 2080Ti, I'll probably get it even at 3090 prices. Maybe a 3090Ti
 
#25 ·
Really the A100? didnt know that.....

Actually that sounds like what they will actually do. I highly doubt they would make a TI variant on the same process they are using now, they would be insane to go over the current 350W TDP. 350W is too much for any modern day video card.... We will probably see these variants after the Big Navi release. Exciting year for graphics technology :). We saw some major software and hardware developments this year. Not the same old boring performance jump like usually generations. Surprising with covid around and all.
 
#29 ·
Because that's 350w of heat that has to be dealt with. RTX 3090 does this by using a 3 slot cooler and the biggest GPU ever produced for desktop. Too big.
 
#27 ·
AMD will start the move to 5nm in 2021-2022. I can see Nvidia pull another 12nm transition. Also think a refresh is a terrible idea for new process unless they also plan to use 7nm for the next architecture.
 
#32 ·
The only way this will be true is if they maintained parallel design/development on Samsung 8nm and TSMC 7nm. You cannot just copy/paste Samsung 8nm (derivative of 10nm) design over to TSMC 7nm without major redesign work.

Nvidia did say in the past they would use TSMC 7nm for most Ampere. Theoretically they could have used Samsung 8nm as a stop-gap because TSMC was not going to work within the timeline they wanted to release. If that ends up being the case, RIP launch buyers.

That said, taking this one with a grain of salt.
 
#34 ·
Yeah. It's more likely an titan variant of the A100 platform will be on it. The A100 is already on TSMC 7nm. I hope Nvidia learned its lesson and bought up TSMC 5nm space for its 2022 launch of Hopper. It's suppose to be complete re-thinking of its GPU. 2022 is really going to be interesting year with AMD continued strong development products coming, Intel finally moving away from Core Architecture (though they still haven't solved ringbus successor).

I do wonder with Intel internal power struggle over (Bob Swan getting taking control) could we see Jim Keller Intel gig continue longer than the 6month consultation. Intel going fabless till it solves its own fab issues hopefully can bring more competition to the industry.
 
#36 ·
Rumors originally stated that Nvidia had a design for 7nm TSMC they were just haggling with TSMC to bring the price down and using Samsung's fabs to put pressure on them (I'm guessing they thought they could get away with it due to Huawei being kicked out), which resulted in failure which forced Nvidia to go with Samsung due to all of TSMC's capacity being booked.

So I wouldnt be suprised if they re-released the cards with just a (few more cores and a)Ti suffix slapped on.
 
#37 ·
#41 ·
So, supposedly, nVidia are canceling the cards that they would eventually release to supersede the cards they haven't really released yet?

Cool.
 
#42 ·
They built up some really good Nvidia vibes with the 3080 announcement and had a lot of that turn sour once people couldn't buy one. If they offer another sku it will reduce the minimal amount of 10gb 3080s available and everyone will see it as nvidia twisting arms to get people to buy the more expensive model.

It would do them no good at all to send out this awesome 20gb model to reviewers (in response to 6900) if people can't buy them. If people can buy them (20gb model) it will piss off everyone that wants the 10gb model, which is probably most people (including myself).
 
#43 ·
20gb 3080 will get bashed and poor reviews. The 10gb of vram will make zero difference in gaming benchmarks, only thing it will do is make a more expensive sku with no performance benfit.


Nvidia probaly figured they can not afford any more negativity at this point, so best to just axe the idea.
 
This is an older thread, you may not receive a response, and could be reviving an old thread. Please consider creating a new thread.
Top