Discussion Leading Edge Foundry Node advances (TSMC, Samsung Foundry, Intel)

[DisEnchantment]

TSMC's N7 EUV is now in its second year of production and N5 is contributing to revenue for TSMC this quarter. N3 is scheduled for 2022 and I believe they have a good chance to reach that target.


N7 performance is more or less understood.


This year and next year TSMC is mainly increasing capacity to meet demands.

TSMC Radically Boosts CapEx to Expand Production Capacities, To Reach $14B For 2019

www.anandtech.com

For Samsung the nodes are basically the same from 7LPP to 4 LPE, they just add incremental scaling boosters while the bulk of the tech is the same.

Samsung is already shipping 7LPP and will ship 6LPP in H2. Hopefully they fix any issues if at all.
They have two more intermediate nodes in between before going to 3GAE, most likely 5LPE will ship next year but for 4LPE it will probably be back to back with 3GAA since 3GAA is a parallel development with 7LPP enhancements.

Samsung’s Aggressive EUV Plans: 6nm Production in H2, 5nm & 4nm On Track

www.anandtech.com

Samsung's 3GAA will go for HVM in 2022 most likely, similar timeframe to TSMC's N3.
There are major differences in how the transistor will be fabricated due to the GAA but density for sure Samsung will be behind N3.
But there might be advantages for Samsung with regards to power and performance, so it may be better suited for some applications.
But for now we don't know how much of this is true and we can only rely on the marketing material.

This year there should be a lot more available wafers due to lack of demand from Smartphone vendors and increased capacity from TSMC and Samsung.
Lots of SoCs which dont need to be top end will be fabbed with N7 or 7LPP/6LPP instead of N5, so there will be lots of wafers around.

Most of the current 7nm designs are far from the advertized density from TSMC and Samsung. There is still potential for density increase compared to currently shipping products.
N5 is going to be the leading foundry node for the next couple of years.

For a lot of fabless companies out there, the processes and capacity available are quite good.

 

[DisEnchantment]

GAA Is Ready for Customers’ Adoption – 3nm MP in 2022, 2nm in 2025
With its enhanced power, performance and flexible design capability, Samsung’s unique GAA technology, Multi-Bridge-Channel FET (MBCFETTM), is essential for continuing process migration. Samsung’s first 3nm GAA process node utilizing MBCFET will allow up to 35 percent decrease in area, 30 percent higher performance or 50 percent lower power consumption compared to the 5nm process. In addition to power, performance and area (PPA) improvements, as its process maturity has increased, 3nm’s logic yield is approaching a similar level to the 4nm process, which is currently in mass production.

3nm’s logic yield is approaching a similar level to the 4nm

Samsung is scheduled to start producing its customers’ first 3nm-based chip designs in the first half of 2022

Samsung Foundry Innovations Power the Future of Big Data, AI/ML and Smart, Connected Devices

Details were disclosed at the company’s 5th Annual Samsung Foundry Forum

news.samsung.com

Its official, 3GAE is still alive. Seems they have been very secretive about GAA

• Density would be comparable to N5 (my guess)
• But the performance and efficiency is much improved compared to 5LPE.
• Yield similar to 4LPE, not bad on first gen GAA
• Samsung is scheduled to start producing its customers’ first 3nm-based chip designs in the first half of 2022

I guess 4LPP and 3GAE will be the lead processes in 2022.
50% power efficiency gain from 5LPE(P?) --> 3GAE is quite impressive.
30% perf gain from 5LPE(P?) --> 3GAE is equally impressive.
Density is OK, important but even TSMC's figure were very ideal case only and not achievable.
These are much closer to the 2019 figures.
I guess the previous Marketing PR was probably a red herring.
Waiting for David Schor to dig deeper.
Really looking forward to see chips on this.

 

[Arkaign]

Samsung is doing really impressive things considering the difference in history vs TSMC/Intel, and especially GF. GF is a basketcase overall.

It's also an era where "X"nm is becoming nearly meaningless. So many different areas of a finished die have wildly differing measurements, something like transistor count per square mm makes slightly more sense but of course is still imperfect and incomplete.

After watching a good bit of Derbauer and reading various engineering white papers related to EUV, Finfet, yadda etc, it becomes a case where predicting exact performance, scalability, yield rate, progress from risk to bulk gens, etc are just impossible to nail down even without taking potential delays and tooling tweaks into account. Proof is in the (final) pudding so to speak. Us old timers of course remember things like 90nm being unexpectedly terrible, 22nm offering shockingly little improvement overall vs 32nm, etc shouldn't find this too surprising as these newer processes are introduced. Until you have a finished product to test, it's never a sure thing.

 

[DrMrLordX]

I guess 4LPP and 3GAE will be the lead processes in 2022.

When will 3GAP be available?

 

[DisEnchantment]

When will 3GAP be available?

3GAP 2023, then 2GAP 2025
Since these are official data in the SFF event I would trust these more rather some marketing PR blurb.

 

[Ajay]

This is all great news for Samsung. The question is whether they will gain any large customers outside Qualcomm. They blew their best opportunity for a large high performance customer with Nvidia. I don’t know what Samsung can do to lure back these sort of profitable contracts.

 

[Doug S]

This is all great news for Samsung. The question is whether they will gain any large customers outside Qualcomm. They blew their best opportunity for a large high performance customer with Nvidia. I don’t know what Samsung can do to lure back these sort of profitable contracts.

So long as TSMC's most advanced processes remain fully booked Samsung may not have to do anything other than be able to offer less lead time.

 

[BorisTheBlade82]

Samsung is doing really impressive things considering the difference in history vs TSMC/Intel, and especially GF.

Which differences do you mean?
Also I have to say that Samsung's announcements always have been impressive in recent years. It's the execution where they failed - especially in comparison to TSMC. I am hoping for them to improve as more Foundry competition is in dire need.

 

[DrMrLordX]

They blew their best opportunity for a large high performance customer with Nvidia.

That's a little surprising since NV probably could have gotten all the 3GAE they wanted which should outperform TSMC N7.

 

[Arkaign]

Which differences do you mean?
Also I have to say that Samsung's announcements always have been impressive in recent years. It's the execution where they failed - especially in comparison to TSMC. I am hoping for them to improve as more Foundry competition is in dire need.

Ah, I could see how that could be misconstrued.

What I mean is that TSMC, GF, and of course Intel fabrication of leading/near leading edge silicon goes back far longer than Samsung, which began relatively late (only launching the division in 2005, which of course takes some time to tool up, gather employees and institutional knowledge, pipeline of dev work and customers, risk projects, yadda).

In general I think it would be agreed that they've surpassed GF, are relatively competitive with Intel, and are not catastrophically behind TSMC. 🤷🤔

 

[BorisTheBlade82]

@Arkaign
Ah, understood and yes, I agree with you.

 

[DrMrLordX]

In general I think it would be agreed that they've surpassed GF, are relatively competitive with Intel, and are not catastrophically behind TSMC. 🤷🤔

If they have not surpassed Intel by now, they will have done so by the time 3GAE goes into full production.

 

[Ajay]

That's a little surprising since NV probably could have gotten all the 3GAE they wanted which should outperform TSMC N7.

Yeah, don't entirely get that. Nvidia must have lost confidence if SS by 2019.

 

[darkswordsman17]

Yeah, don't entirely get that. Nvidia must have lost confidence if SS by 2019.

I think its been reported that SS's 7nm still is offering poor yields?

That's why I think the move to do the licensing deal with Samsung was a smart one by AMD. It would give them some access to Samsung's process without having to commit themselves to it. And Samsung is going to produce those chips regardless (so AMD can also see how it changes over time to see if Samsung is improving any, plus the future versions will almost certainly be produced on the next newest node, so they'll keep having a bit of an idea).

Which, if Samsung actually gets it right (I do wonder if after not keeping up with TSMC on say 7nm, they didn't push most of the resources to trying to get there with 5nm, and then even more quickly tried to push for the next one to try and get there before TSMC in order to build hype), someone might reap some significant benefits, but there's a reason why TSMC is booked up and Intel, AMD, NVidia, Apple, Qualcomm, and everyone else keep passing on Samsung. Qualcomm waffles to and from them a lot (and I wonder how much of that is because Samsung is one of if not their largest customer for SoCs, otherwise they might would ditch them for good, and they did announce they were evaluating Intel's fab so it seems like there's not a ton of loyalty there although some of that might be because Samsung seems like they might look to move on from using Qualcomm as much).

Samsung passed GF years ago, since GF couldn't even get 14nm working right and licensed a working version from Samsung. That's already been what 5-6 years now?

 

[moinmoin]

What I mean is that TSMC, GF, and of course Intel fabrication of leading/near leading edge silicon goes back far longer than Samsung, which began relatively late (only launching the division in 2005, which of course takes some time to tool up, gather employees and institutional knowledge, pipeline of dev work and customers, risk projects, yadda).

That's wrong. Samsung is in the semiconductor business since 1974 actually. What changed in 2005 is that they opened up their production capability to external fabless companies.

 

[Saylick]

That's a little surprising since NV probably could have gotten all the 3GAE they wanted which should outperform TSMC N7.

I suspect the yields on 3GAE aren't good enough to manufacture monolithic 500mm2+ dies. Samsung's 8nm node is basically a refined version of their 10nm node, so it was very mature by the time Ampere was to launch. If I recall correctly, Nvidia originally wanted to use Samsung's 7nm node for Ampere, which is a true new node vs. their 10nm node. Unfortunately, Samsung's 7nm node wasn't mature enough at the time, so Nvidia bit the bullet and resigned to using their 8nm node.

While SS's 7nm node family is probably yielding much better than before, I suspect Nvidia is feeling the pressure from AMD and they know that if they use SS's 7nm family, it would have performance characteristics worse than TSMC's N5 node. JHH's big fat ego would never let them squander a performance crown, so they too jumped onboard the TSMC N5 train to give them the best chance of heading off AMD and RDNA 3.

 

[Ajay]

I think its been reported that SS's 7nm still is offering poor yields?

8N - but yeah, that would be bad.

I suspect the yields on 3GAE aren't good enough to manufacture monolithic 500mm2+ dies. Samsung's 8nm node is basically a refined version of their 10nm node, so it was very mature by the time Ampere was to launch. If I recall correctly, Nvidia originally wanted to use Samsung's 7nm node for Ampere, which is a true new node vs. their 10nm node. Unfortunately, Samsung's 7nm node wasn't mature enough at the time, so Nvidia bit the bullet and resigned to using their 8nm node.

Well, wasn't 3GAP supposed to be out before 3GAE originally. 3N GAA, unless SS has screwed it up, offers some really big advances over FinFETs. Nanosheets/wires are apparently more amenable to 'tuning'. Tuning out parasitics, and tuning for perf/watt. Add more sheets vertically stacked - an the drive currents can go up ( to some limit defined by the process). Some cools stuff, still, features are pretty small - apparently fabbing isn't terribly different from FinFETs, so the same equipment used for smaller FinFET nodes doesn't need to be changed out. On top of that, high NA EUV machines are shipping, or will do so very soon (for > 3N nodes).

 

[Saylick]

Well, wasn't 3GAP supposed to be out before 3GAE originally. 3N GAA, unless SS has screwed it up, offers some really big advances over FinFETs. Nanosheets/wires are apparently more amenable to 'tuning'. Tuning out parasitics, and tuning for perf/watt. Add more sheets vertically stacked - an the drive currents can go up ( to some limit defined by the process). Some cools stuff, still, features are pretty small - apparently fabbing isn't terribly different from FinFETs, so the same equipment used for smaller FinFET nodes doesn't need to be changed out. On top of that, high NA EUV machines are shipping, or will do so very soon (for > 3N nodes).

According the roadmap image in my last post, 3GAP comes after 3GAE, which is consistent with Samsung's nomenclature for their previous nodes, i.e. E (early) nodes come before P (plus) nodes.

I agree that nanosheets are a significant breakthrough versus FinFET because you can tune the dimensions of the nanosheets far more finely than having to select the number of discretized fins. However, I think it is too much of a stretch to say that fabbing them is not that much different than fabbing FinFETs. Otherwise, I think TSMC would have already moved onto nanosheets by now. I don't have a background in semiconductor manufacturing, but I trust that if TSMC is delaying their rollout of a nanosheet node, it is for good reason and I am inclined to believe them because they've made the right moves for the last half-decade. Samsung chose to be aggressive with implementing nanosheets because they have to if they want to catch up to TSMC. TSMC can afford to be conservative with their tried-and-true strategy of slow-and-steady-wins-the-race, but it is evident that implementing nanosheets is not like FinFETs. Samsung has already publicly announced a delay on their GAA node, so it's clear that it ain't an easy lift.

To wrap up, Nvidia can try to develop a future product on SS's GAA node, but that would be risky if they don't want to secede any market share to their competitors. If Samsung flubs again like they did with 7nm, Nvidia don't want to be caught having to backport a design to an inferior node. Hence, JHH cannot afford to take chances with the foundry and has selected TSMC for their next generation products.

 

[DisEnchantment]

We can bet shareholders are happy NV went with Samsung for current gen, imagine being on N7. They wouldn't be anywhere near to half those reported revenue. For NV, Samsung 8N is a major success.
My guess is that Qualcomm went to Intel for the Government contract and that is understandable because the product requires everything made in America.
And for comms and baseband chips you definitely want that. In the age of soft baseband logic, you better be sure your high tech chips meant for a future bomber are not tampered.
But Qualcomm will have to work with Samsung to fab their mobile SoCs, otherwise they have to get in line for N5. Waiting for N5/3 means their 20 Billion revenue could vaporize overnight.
We can expect NV to fab at TSMC and Samsung otherwise they can say goodbye to their impressive revenues. Auto is anyway on Samsung

With GAA, so far only Samsung has a a design kit certified at both Synopsys and Cadence. Nothing so far from Intel or TSMC.
For Samsung's recent nodes, 3GAE brings the biggest jump in perf and efficiency over a previous node (as advertized), so once we see GAA silicon, it should be interesting.

 

[Ajay]

@Saylick
This article on GAA is pretty good in terms of being a quick design and manufacturing overview: https://semiengineering.com/impact-of-gaa-transistors-at-3-2nm/

 

[Arkaign]

That's wrong. Samsung is in the semiconductor business since 1974 actually. What changed in 2005 is that they opened up their production capability to external fabless companies.

Sort of. Re-read what I wrote, it wasn't that Samsung had no semiconductor, it was just smaller scale/more basic tech. Dram and small die stuff ala much like a lot of other entities out there.

What changed in 2005 was their huge expansion into the track to do large wafer on competitive process ala TSMC, GF, Intel.

A Detailed History of Samsung Semiconductor - Semiwiki

From our book

semiwiki.com

From : "Another important part in this family was the S3C2443 in 2007, running at 533 MHz on 0.13 micron in a 400-pin package, which found design wins in portable navigation devices like the Asus R300 and the LG LN800."

130nm small die 😶

To : " Pushing its process further, Samsung opened its 300mm S1 line in Giheung in mid-2005, establishing production of DRAM on a 90nm process by the end of the year. To develop the next 65nm process node, estimated to cost $5B or more, Samsung allied with Chartered Semiconductor and IBM in the Common Platform initiative. IBM piloted a 65nm process at its Fishkill, New York facility and Samsung was preparing to roll it out at Giheung in early 2006."

This was the roots of being able to scale beyond niche and tertiary products to flagship risk/large die ICs. It took huge investments and cooperation with outside specialists, but has paid off pretty well. From effective processors for mobile, now all the way to fairly huge Nvidia GPU dies. It hasn't been all smooth sailing, but it definitely shows how expensive and involved it is to even play in the same world as the other big foundries.

It's fascinating to see how many active foundries are out there :

https://en.m.wikipedia.org/wiki/List_of_semiconductor_fabrication_plants

Yet how few could produce orders to say even Zen2 or 12nm Nvidia dies (on competitive process, scale, and yield).

 

[moinmoin]

Sort of. Re-read what I wrote, it wasn't that Samsung had no semiconductor, it was just smaller scale/more basic tech. Dram and small die stuff ala much like a lot of other entities out there.

That's still wrong. That what you call "smaller scale/more basic tech" is still the biggest customer of Samsung Semiconductor. Samsung's memory and storage business is still the biggest customer of Samsung's foundries.

What you can pin down to 2005 is the epiphany that bleeding edge node development is not sustainable if it's only used for in-house products which can and often will lack the necessary scale. Samsung got that right one and a half decades earlier than Intel.

 

[DrMrLordX]

I suspect the yields on 3GAE aren't good enough to manufacture monolithic 500mm2+ dies. Samsung's 8nm node is basically a refined version of their 10nm node, so it was very mature by the time Ampere was to launch. If I recall correctly, Nvidia originally wanted to use Samsung's 7nm node for Ampere, which is a true new node vs. their 10nm node. Unfortunately, Samsung's 7nm node wasn't mature enough at the time, so Nvidia bit the bullet and resigned to using their 8nm node.

Samsung's 7LPP turned out to be quite the disaster on Snapdragon 888. It's serviceable but it really isn't any better than N7.

While SS's 7nm node family is probably yielding much better than before, I suspect Nvidia is feeling the pressure from AMD and they know that if they use SS's 7nm family, it would have performance characteristics worse than TSMC's N5 node. JHH's big fat ego would never let them squander a performance crown, so they too jumped onboard the TSMC N5 train to give them the best chance of heading off AMD and RDNA 3.

Pretty sure NV is going to jump from Samsung 8LPP to N7, not N5. At least for the Ampere refresh. Maybe NV will get N5 wafers later.

 

[Mopetar]

If they were going to go back to TSMC it was likely a decision they made a while back because it's probably next to impossible to get enough wafers without paying through the roof at this point.

If Samsung can give NVidia better volume they'd be foolish to ditch them for slightly better performance when they're going to sell anything they can make and that's unlikely to change for the foreseeable future.

AMD might want to consider moving something to Samsung as well if there's more wafer availability there. Not all of their products need to be on TSMC and they're pretty wafer constrained as well.

 

[LightningZ71]

It's likely that both GPU vendors will split their products, going forward, between scaled down trailing node products at the low end, and big, bleeding edge nodes on the higher end.

 

[Doug S]

My take is that when it comes to trying new things TSMC is more conservative and Samsung is more aggressive - I guess based on "when you're in second place you need to be more aggressive if you want to move into first".

Samsung went all-in on EUV for critical layers in 7nm, while TSMC stuck with DUV and only added EUV steps later for N7+ - and even that wasn't going all-in since they also offered the all DUV N7P alongside it. That allowed customers like Apple who can't afford yield killing mistakes to have a stable node to produce SoCs for their most important product. Sure, if TSMC had gone EUV, and done it right, their 7nm could have been better than it was. But if things didn't go right it could have been a disaster (see Intel's 10nm rollout)

I see TSMC's decision to stick with FinFET in N3 and wait on GAA until N2, while Samsung goes GAA for their 3nm generation, as just another example of this. If Samsung gets it right they may get a leg up on TSMC. But just like their issues rolling out 7nm due to its reliance on EUV the schedule slips on 3nm seem to indicate a gamble that has failed to pay off.

Also consider that Apple has process engineers who work alongside TSMC's staff in the development of new processes. It isn't clear in what form this cooperation takes place. Is Apple just wanting to learn the characteristics of upcoming processes as early in advance to feed back to their architects? Or are they actually helping make the call for when to integrate EUV or GAA? As TSMC's largest customer and one who pays years in advance and thus has had a lot to do with where TSMC is today, that's quite possible.