Support NeoGAF

[winjer]

NVIDIA Is Now Rumored To Switch Towards Samsung Foundry For 2nm Process, Ditching TSMC Due To High Costs

NVIDIA is rumored to ditch TSMC's 2nm process, and instead turn towards Samsung Foundry, given the high costs associated.

wccftech.com

Before we dive into the report, it is essential to take this rumor with a grain of salt. Right now, mainstream 2nm suppliers include TSMC, Samsung, and the Japanese firm Rapidus, with the latter two companies "apparently" being behind the Taiwan giant in terms of yield rates. While this trend has been consistent with previous nodes as well, it seems like TSMC will experience a greater rivalry this time, given that its key clients are now focused on diversifying the 2nm supply, adding in new partners.

In particular, it is claimed that NVIDIA is exploring the option of getting 2nm from Samsung Foundry; however, the firm is reluctant for now, given Samsung's past track record with clients like Qualcomm. Apart from Team Green, we recently reported on how Qualcomm plans to adopt a "dual-source" strategy, including both TSMC and Samsung in its 2nm supply chain, so it won't be wrong to say that the next "big process" will likely have more competitors, rather than TSMC alone.

The adoption of 2nm looks slowed for now, given that Apple has shifted its plans to 2026, and with NVIDIA and Qualcomm exploring other options, it seems like integration in consumer products will take time. However, given that Samsung's 2nm process fails to impress the markets, companies will have no other option and will ultimately revert to TSMC.

I doubt Nvidia would use Samsung for their top AI products, for enterprises. Simply because they need the best process nodes in the world to maintain their advantage, but also because these products have such high profit margins, that Nvidia can afford TSMC's rates.
But I can see Nvidia using Samsung for cheaper products, where price is more of a concern. Maybe for mid and low end GPUs, embedded systems, chips for automobiles, etc.

 

[LordOcidax]

2nm, that’s crazy. That’s smaller than a DNA molecule… next, 0.5nm, the size of an Atom.

 

[winjer]

2nm, that’s crazy. That’s smaller than a DNA molecule… next, 0.5nm, the size of an Atom.

These measurements are not realistic. It's just marketing.
It used to be that it would refer to the length of the gate. But that has stopped happening nearly a decade ago.

 

[asdasdasdbb]

I doubt Nvidia would use Samsung for their top AI products, for enterprises. Simply because they need the best process nodes in the world to maintain their advantage, but also because these products have such high profit margins, that Nvidia can afford TSMC's rates.
But I can see Nvidia using Samsung for cheaper products, where price is more of a concern. Maybe for mid and low end GPUs, embedded systems, chips for automobiles, etc.

Power consumption being the main concern is why you would use N2 over anything else (including just using N3X as an alternative as well).

My guess is Rubin Gaming uses N3X. Using a Samsung node for the lower end products should remain an option tho.

 

[V1LÆM]

my first computer was around 100nm.. maybe even closer to 200nm (it was a windows 95 machine). my first mac was 40 something nm. crazy how far we've come.

what happens when we get to <1 nm ? new marketing name?

 

[LordOcidax]

These measurements are not realistic. It's just marketing.
It used to be that it would refer to the length of the gate. But that has stopped happening nearly a decade ago.

I always thought that was the size of the transistors in the chip. Still crazy technology.

 

[V1LÆM]

I always thought that was the size of the transistors in the chip. Still crazy technology.

it got more difficult to reduce size and sticking with the physical size wasn't good for marketing. intel were stuck on 14nm for a long time. there was a joke that it was 14nm++++

also with us approaching 2nm it doesn't leave long before we're at <1nm. gonna need to come up with a new name to make things sound better. we might get 1.8, 1.6, 1.4, etc to kick the can down the road a bit more.

edit:

also 3nm chips aren't all 3nm parts. some parts are 20-50nm in size.

my CPU is 5nm but that's just the cores. the IO is 6nm.

 

[winjer]

it got more difficult to reduce size and sticking with the physical size wasn't good for marketing. intel were stuck on 14nm for a long time. there was a joke that it was 14nm++++

also with us approaching 2nm it doesn't leave long before we're at <1nm. gonna need to come up with a new name to make things sound better. we might get 1.8, 1.6, 1.4, etc to kick the can down the road a bit more.

It's called Angstroms. For example, Intel already markets it's next generation node as 18A.

 

[V1LÆM]

It's called Angstroms. For example, Intel already markets it's next generation node as 18A.

i always learn something new from your posts

 

[Trogdor1123]

These measurements are not realistic. It's just marketing.
It used to be that it would refer to the length of the gate. But that has stopped happening nearly a decade ago.

I used to be amazed by these two until it was explained to me. It’s still utterly mind boggling but the numbers are really not what we thought they are.

 

[vaibhavpisal]

what happens when we get to <1 nm ? new marketing name?

They will have to find a replacement material to silicon.

 

[Katatonic]

NVM.

 

[BennyBlanco]

i always learn something new from your posts

winjer really is the man. Never even seen him argue with anyone.

 

[StereoVsn]

I used to be amazed by these two until it was explained to me. It’s still utterly mind boggling but the numbers are really not what we thought they are.

Yeah, let’s not let the exact measurement mask the sheer technical wonder that is the modern silicon. It’s really crazy and equivalent or even higher difficulty be rocket science.

I like watching Asinometry YouTube on the subject and the complexity is crazy. Whether it’s 5nm, 4nm, 3nm or other nodes the tech behind EUV lithography is just crazy.

Edit: It would be good to have another competitor for newer processes if Samsung can get yields up. gaming GPUs could be done on Samsung nodes in that case if not the really expensive AI targeted modules.

 

[Dr.D00p]

So, Nvidia are crying about how expensive it is becoming to make their GPUs.

Try buying them, Motherfuckers...

 

[daxgame]

let's see how long it takes now for someone to report that Switch 2 uses Samsung's 2nm

 

[diffusionx]

my first computer was around 100nm.. maybe even closer to 200nm (it was a windows 95 machine). my first mac was 40 something nm. crazy how far we've come.

what happens when we get to <1 nm ? new marketing name?

In the Windows 95 era, it was like 350nm-500nm. They didn't get down to 100nm until like, 2005.

Like others have said, this is a number that used to have a specific meaning, but no longer does, because they can't really get it down to the numbers that they say they are at. I know people goof about Intel's 14nm++++++++++++++++++++++++ process but it's more descriptive and accurate than what TSMC claims.

 

[Buggy Loop]

winjer really is the man. Never even seen him argue with anyone.

I mean I still like winjer but he argued with me so many fucking times

 

[nkarafo]

also with us approaching 2nm it doesn't leave long before we're at <1nm. gonna need to come up with a new name to make things sound better. we might get 1.8, 1.6, 1.4, etc to kick the can down the road a bit more.

We are very close to the point where it won't be physically possible to reduce the sizes further. I think we are almost already there?

There needs to be a completely new and groundbreaking technology from now on, something as big as the jump from tubes to transistors, if we want to see computers advance any further in the near future.

 

[asdasdasdbb]

We are very close to the point where it won't be physically possible to reduce the sizes further. I think we are almost already there?

Probably hit economic practicality first.

 

[Setsuna Mudou]

2nm, that’s crazy. That’s smaller than a DNA molecule… next, 0.5nm, the size of an Atom.

Yeah it's currently only a theoretical measure of performance if it was possible to reduce the size of the gates to such sizes. Actually doing it would kill the logic because you'd have a gate of lower size than the wavelegth of the electron, thus not able to produce a 0.

edit: nvm, winjer already covered it in the respose

 

[GHG]

We are very close to the point where it won't be physically possible to reduce the sizes further. I think we are almost already there?

There needs to be a completely new and groundbreaking technology from now on, something as big as the jump from tubes to transistors, if we want to see computers advance any further in the near future.

Graphene is the material that is furthest ahead in terms of research for the next step, but it would mean a rough transitional period.

Researchers Create First Functional Semiconductor Made From Graphene | Research

Researchers at the Georgia Institute of Technology have created the world’s first functional semiconductor made from graphene, a single sheet of carbon atoms held together by the strongest bonds known. Semiconductors, which are materials that conduct electricity under specific conditions, are...

research.gatech.edu

 

[LordOfChaos]

Nvidia so far ahead they log into their smurfing account with a cheaper Sammy node sometimes and still stay ahead

Come on, Intel

 

[Buggy Loop]

Maybe Nvidia’s AI litho will help Samsung a lot? They must have ironed out a few problems with yields if Nintendo is gonna mass produce 100M chips there.

 

[TheRedRiders]

I don't blame Nvidia to some extent, TSMC's fabs are starting to get absurdly pricy.

It will be a shame as i do think they'll miss out on certain efficiency and performance gains but then again Ampere was an excellent architecture, and did a solid job against RDNA 2 even though they were at a node disadvantage (Samsung's 8nm vs TSCM's 7nm).

 

[IFireflyl]

I mean I still like winjer but he argued with me so many fucking times

I think BennyBlanco meant winjer doesn't argue with people who matter.

Spoiler

 

[Beechos]

let's see how long it takes now for someone to report that Switch 2 uses Samsung's 2nm

New switch 2 secret sauce! How bold of them to sneak it into the headline.

NVIDIA Is Now Rumored To Switch Towards Samsung Foundry For 2nm Process.​

C'mon follow the breadcrumbs. Its all part of the deal for using Samsung groundbreaking nand storage for the switch 2.

 

[Gameplay Gods Bless]

NVIDIA Is Now Rumored To Switch Towards Samsung Foundry For 2nm Process, Ditching TSMC Due To High Costs

NVIDIA is rumored to ditch TSMC's 2nm process, and instead turn towards Samsung Foundry, given the high costs associated.

wccftech.com

I doubt Nvidia would use Samsung for their top AI products, for enterprises. Simply because they need the best process nodes in the world to maintain their advantage, but also because these products have such high profit margins, that Nvidia can afford TSMC's rates.
But I can see Nvidia using Samsung for cheaper products, where price is more of a concern. Maybe for mid and low end GPUs, embedded systems, chips for automobiles, etc.

Nvidia doesn't need the best process in the world. Case in point their next gen GPUs both for data centers and client are using an old 5nm family process despite next gen 3N family processes available and used by the likes of Intel and Apple.

 

[Gameplay Gods Bless]

my first computer was around 100nm.. maybe even closer to 200nm (it was a windows 95 machine). my first mac was 40 something nm. crazy how far we've come.

what happens when we get to <1 nm ? new marketing name?

Intel is already moved on to using Angstroms such as their upcoming 18A process.

 

[Haint]

Nvidia so far ahead they log into their smurfing account with a cheaper Sammy node sometimes and still stay ahead

Come on, Intel

In all seriousness, I have long suspected Nvidia's new products are actually at least a generation older/slower than what they could release if someone actually caught them off guard. No need to try though when their closest competition has basically been a full generation behind for over a decade at this point, they just intentionally handicap themselves to keep the break in case of emergency trap card in their back pocket.

 

[Hudo]

Almost the size of my penis.

 

[RCX]

And all to give you graphics that are partially better than a ps5 for only four times the price!

What a savings!!

 

[lh032]

nano process ,galactic price.

 

[LordOfChaos]

In all seriousness, I have long suspected Nvidia's new products are actually at least a generation older/slower than what they could release if someone actually caught them off guard. No need to try though when their closest competition has basically been a full generation behind for over a decade at this point, they just intentionally handicap themselves to keep the break in case of emergency trap card in their back pocket.

My comment was basically in all seriousness lol. A few generations back they basically shifted all the dies down a tier for what you pay because they were so far ahead of AMD, even used a significantly lacking Samsung node for a generation to pocket some extra money, and stayed completely dominant. They are 100% smurfing right now and I'm almost giving up on AMD getting them to get serious, maybe Intel will in another generation or two.

Maybe Nvidia’s AI litho will help Samsung a lot? They must have ironed out a few problems with yields if Nintendo is gonna mass produce 100M chips there.

Nintendo isn't too much related because they're using an older 8nm node no one else really used, again which is similar to the Switch and their withered tech ethos

 

[Buggy Loop]

Nintendo isn't too much related because they're using an older 8nm node no one else really used, again which is similar to the Switch and their withered tech ethos

Again., nobody knows the node, nobody.

 

[LordOfChaos]

Again., nobody knows the node, nobody.

I mean a lot of people at Nintendo, Samsung, and Nvidia know

I remember a lot of nah it's not that node at the NX speculation and then it was the node

It might be wrong but it might also be right and sounds about like what Nintendo has done

 

[LordOfChaos]

Again., nobody knows the node, nobody.

 

[Buggy Loop]

Yea, they don't know more than any forum speculation.

Rich doesn't understand or hasn't counted that T234 SMs account for 10B transistors. even at 75% cutoff from the rumoured T239 specs, you're 7.5B transistors. Keeping same node as Samsung 8nm of Ampere and keeping the same density means you have 1.5B transistor budget left for everything else that makes the APU, including the rumoured Ada backport features. Unlikely.

Samsung's 7nm line witch 5LPH is more likely. Nvidia never goes into full density of a node. A 60 MTr/mm^2 would sit well with a 12B transistor APU to 200mm^2.