TSMC 4nm vs Samsung 4nm

[prime007]

The tech press covered Qualcomm's "Snapdragon Night" which introduced the Snapdragon 8+ Gen 1. What maybe interesting to us is that we can (kinda) compare TSMC 4nm vs Samsung Foundry 4nm process using the Snapdragon 8 Gen 1 chip. This hasn't happened since the Apple A9 processor back in 2019 and those metrics are the basis used for Scotten's wonderful foundry node comparison articles!

Here are a few interesting article snippets below...

"[D]iving into the specifications, the new Snapdragon 8+ Gen 1 is essentially the original Snapdragon 8 Gen 1 ported over from Samsung’s 4nm line over to one of TSMC’s 4nm lines. Under more normal circumstances, this kind of a shift would likely be unremarkable – or at most, an amusing exercise in looking for edge cases – however for Qualcomm’s flagship SoC, the matter is more significant.
...
While official sources and statements on the quality of Samsung’s 4nm process are few and far between, unofficially, it’s become clear that Samsung’s 4nm process hasn’t lived up to expectations. This has caused a cascading impact on the chips made on the process node, leading to the original Snapdragon 8 Gen 1 developing an affinity for power consumption, and Samsung’s own Exynos 2200 not faring any better. Conversely, by all accounts TSMC's N4 process is looking great, with the optically shrunk node building off of TSMC’s already successful and very performant 5nm technologies." - Anandtech

"The company said the chip will provide "10 percent faster CPU performance," thanks to a 200 MHz peak CPU boost (up to 3.2 GHz now) and a 10 percent faster GPU. The real shocker is a "30 percent improved power efficiency" claim for the CPU and GPU.

For the Snapdragon 8 Gen 1 Plus, Qualcomm is moving the chip from Samsung Foundry to TSMC, which is apparently where the power improvements are coming from. That's a serious slam against Samsung's 4 nm process versus TSMC's 4 nm process, but it lines up with earlier reports of troubles at Samsung Foundry."

Swapping foundries as part of a mid-cycle upgrade is not normal, and it seems that Qualcomm has a bit of a salvage operation on its hands with the Snapdragon 8 Gen 1. The chip has not fared very well in the real world, with the CPU regularly turning in lower benchmark scores than 2021's flagship Snapdragon 888. - Ars Technica

Sources:

Qualcomm Announces Snapdragon 8+ Gen 1: Moving to TSMC for More Speed, Lower Power

www.anandtech.com

Qualcomm’s Snapdragon “8+ Gen 1” salvage operation moves the chip to TSMC

After missing performance estimates, Qualcomm is dumping Samsung mid-cycle.

arstechnica.com

 

[benb]

So how do you differentiate 8+ Gen 1 from Mediatek’s version? ARM cores, same node, same performance, same foundry. Does the industry need two high volume identical offerings?
As we approach the event horizon of sameness, one price, determined by TSMC, I ask who is winning? Not Qualcomm or Mediatek. They risk becoming white labels.

Despite all the issues, Samsung Foundry, like AMD, exists to provide an alternative, of some kind, so that competitive strategy doesn’t collapse on itself.

 

[Daniel Nenni]

Despite all the issues, Samsung Foundry, like AMD, exists to provide an alternative, of some kind, so that competitive strategy doesn’t collapse on itself.

True, the NOT TSMC strategy works until you get a technical advantage. AMD rode the NOT Intel bus for many years until they got the upper hand. I think Intel Foundry is bad news for Samsung in the NOT TSMC business. IDM Foundry vs IDM Foundry is a much more even fight and I give the edge to Intel now that they bought Tower, absolutely.

 

[prime007]

So how do you differentiate 8+ Gen 1 from Mediatek’s version? ARM cores, same node, same performance, same foundry. Does the industry need two high volume identical offerings?
As we approach the event horizon of sameness, one price, determined by TSMC, I ask who is winning? Not Qualcomm or Mediatek. They risk becoming white labels.

Despite all the issues, Samsung Foundry, like AMD, exists to provide an alternative, of some kind, so that competitive strategy doesn’t collapse on itself.

For Qualcomm's risk of becoming a white label, their answer may be their $1.4 billion acquisition of Nuvia. I believe Qualcomm indicated that they would use whatever design (either NUVIA's or ARM's) is more performant in the future. I'm curious how Nuvia's PC processors (due in late 2023) will perform vs Apple's forthcoming M2 chips given that Nuvia's CEO, Gerald Williams, was once Apple's chef chip architect and allegedly poached other Apple employees to join him at Nuvia.

During ASML's Q3 2020 earnings call, ASML CEO Peter Wennink stated that he doesn't believe all three of the current leading edge foundries (Intel, Samsung, TSMC) will be able to keep up with the cutting edge due to costs. He indicated that one would eventually have to drop out.

"Now, on the number of customers in the leading edge, as one thing is absolutely certain, things are not getting easier. The next nodes will increase complexity and I think only the very large customers can deal with that. We've seen over the last 20 years a significant reduction in number of leading edge customers and I think it's going to continue with less than a handful. And I'm going to not go to speculate who those are going to be because they are all dear to my heart. And I think they'll have to figure it out who has the best designs, the best production technology, who are the most efficient and the lowest cost. But in the end, it doesn't matter, because in the end what we always talk about is the number of wafers and device and chips that are needed in the digital transformation that need to be made somewhere on the planet.

And that means that I do believe that our large customers are going to be larger and they are going to be dominant in areas of chip production, which is only going to grow. That is why I said at the end of my prepared remarks, our confidence in our 2025 outlook has only grown. As always, ultimately, you could argue also bit as a result of what we've seen as a part of the COVID crisis how important this digital infrastructure is. So, in that sense, yes, I think there will be fewer and fewer customers, but it will be much bigger than they are today." - Peter Wennink, President and CEO of ASML

I believe SK Hynix and Micron are both planning on purchasing EUV equipment (with SK Hynix eyeing a potential entry into the logic fab business) so that may make it harder for everyone (TSMC, Samsung, Intel, SK Hynix and Micron) to get the number of EUV machines they desire from ASML. I imagine this situation will tremendously benefit TSMC by preventing Intel/Samsung from rapidly scaling their production lines.

The cutting-edge foundry space seems like a winner-takes-all market. TSMC is currently winning & profiting handsomely because they have the best process technology. Samsung Foundry currently exists as an alternative but its speculated here that they are currently losing money...and Samsung might be in even more financial pain if Intel Foundry Services replaces Samsung as the preferred alternative. A clue will arrive in 2023 with the launch of Intel's Meteor Lake CPUs (Intel 4).

Qualcomm choose Samsung Foundry over TSMC because of price and paid dearly for it. Samsung was unable to supply Qualcomm with enough chips (poor yields), the Samsung-fabbed chips underperformed and overheated and MediaTek was able to win market share by putting its TSMC-fabbed Dimensity 9000 chips in many high-end phones (OnePlus 10R, OnePlus Nord 3, Vivo X80, Xiaomi 12T, etc..)

 

[prime007]

It looks like Samsung is planning on showcasing a 3nm processor to President Biden during his visit to South Korea.

The visit to the Pyeongtaek factory will be the scene of an important exhibition. Samsung Electronics Vice President Lee Jae-yong will also be present at the visit to exhibit the 3nm GAA (Gate All Around) chips that the company has been working on for a long time.

Although the 3nm GAA has seen a 35 percent reduction in size compared to 5nm chips, it is stated that it will offer 30 percent higher performance. In addition, low power consumption of up to 50 percent is expected. On the other hand, Samsung said that they have also started initiatives for 2nm chips. This chipset is expected to start production as early as 2025.

However, the rumor is that Samsung's 3nm chip currently only yields 10-20% range!

Source(s): https://www.technoscience.fikrikadim.com/2022/05/20/samsungs-3nm-processor-on-display/

Yield on Samsung's 3nm chip production reportedly drops to a 10-20% range

A report released today says that Samsung's yield on the production of its 3nm Gate-all-around chipsets is a poor 10% to 20%, even worse than the 35% yield on its 4nm chips.

www.phonearena.com

 

[hist78]

For Qualcomm's risk of becoming a white label, their answer may be their $1.4 billion acquisition of Nuvia. I believe Qualcomm indicated that they would use whatever design (either NUVIA's or ARM's) is more performant in the future. I'm curious how Nuvia's PC processors (due in late 2023) will perform vs Apple's forthcoming M2 chips given that Nuvia's CEO, Gerald Williams, was once Apple's chef chip architect and allegedly poached other Apple employees to join him at Nuvia.

During ASML's Q3 2020 earnings call, ASML CEO Peter Wennink stated that he doesn't believe all three of the current leading edge foundries (Intel, Samsung, TSMC) will be able to keep up with the cutting edge due to costs. He indicated that one would eventually have to drop out.


I believe SK Hynix and Micron are both planning on purchasing EUV equipment (with SK Hynix eyeing a potential entry into the logic fab business) so that may make it harder for everyone (TSMC, Samsung, Intel, SK Hynix and Micron) to get the number of EUV machines they desire from ASML. I imagine this situation will tremendously benefit TSMC by preventing Intel/Samsung from rapidly scaling their production lines.

The cutting-edge foundry space seems like a winner-takes-all market. TSMC is currently winning & profiting handsomely because they have the best process technology. Samsung Foundry currently exists as an alternative but its speculated here that they are currently losing money...and Samsung might be in even more financial pain if Intel Foundry Services replaces Samsung as the preferred alternative. A clue will arrive in 2023 with the launch of Intel's Meteor Lake CPUs (Intel 4).

Qualcomm choose Samsung Foundry over TSMC because of price and paid dearly for it. Samsung was unable to supply Qualcomm with enough chips (poor yields), the Samsung-fabbed chips underperformed and overheated and MediaTek was able to win market share by putting its TSMC-fabbed Dimensity 9000 chips in many high-end phones (OnePlus 10R, OnePlus Nord 3, Vivo X80, Xiaomi 12T, etc..)

It's better to compare the upcoming Qualcomm processors (possibly based on Nuvia's design) against Intel's or AMD's processors. People buy Mac for reasons and Apple doesn't license Mac OS to any outsiders. Qualcomm is probably eying the Windows market if they really want to go big and go mainstream.

And for Microsoft it's a good hedging to avoid being locked down by Intel.

 

[Fred Chen]

Some teardown (presumably by TechInsights) revealed Snapdragon 8 Gen 1 was not even true 4nm but a so-called 4LPX (very similar to 7/5 nm):

The Snapdragon 8 Gen 1 may be significantly inferior to Samsung's Exynos 2200 - Pokde.Net

Rumors are cropping up that the Snapdragon 8 Gen 1 (or Snapdragon 898) may be significantly worse than the Exynos 2200.

pokde.net

Subsequent Exynos 2200 is true 4LPE:

#samsung #exynos_2200 #4lpe #5lpe #cpp | Yuzo Fukuzaki

We waited for this. Here is Samsung's true 4nm, 4LPE. We saw Qualcomm Snapdragon 8 gen1 has 4LPX and hardly found “4nm” features. Then, our next interst is Samsung Exynos 2200 with 4LPE. Here is quick results. You see visible SDB, 53 nm CPP (Gate pitch, was 54 nm @5LPE) and 198 nm UHD...

www.linkedin.com

 

[Scotten Jones]

The tech press covered Qualcomm's "Snapdragon Night" which introduced the Snapdragon 8+ Gen 1. What maybe interesting to us is that we can (kinda) compare TSMC 4nm vs Samsung Foundry 4nm process using the Snapdragon 8 Gen 1 chip. This hasn't happened since the Apple A9 processor back in 2019 and those metrics are the basis used for Scotten's wonderful foundry node comparison articles!

Here are a few interesting article snippets below...




Sources:

Qualcomm Announces Snapdragon 8+ Gen 1: Moving to TSMC for More Speed, Lower Power

www.anandtech.com

Qualcomm’s Snapdragon “8+ Gen 1” salvage operation moves the chip to TSMC

After missing performance estimates, Qualcomm is dumping Samsung mid-cycle.

arstechnica.com

The 10% performance is a little less of an advantage for TSMC than I would have expected, the 30% better power is actually less too. Samsung 3nm with Horizontal Nanosheets is supposed to close a lot of the power gap versus TSMC but the yields are awful.

 

[Daniel Nenni]

The tech press covered Qualcomm's "Snapdragon Night" which introduced the Snapdragon 8+ Gen 1. What maybe interesting to us is that we can (kinda) compare TSMC 4nm vs Samsung Foundry 4nm process using the Snapdragon 8 Gen 1 chip. This hasn't happened since the Apple A9 processor back in 2019 and those metrics are the basis used for Scotten's wonderful foundry node comparison articles!

The Apple A9 was designed to Samsung 14nm and ported over to TSMC 16nm. Not the same as designing to TSMC 16nm directly. I'm wondering if QCOM ported the design over versus natively designing it? Maybe that explains the performance and power issues. I will check into it.

 

[jorgequinonez]

Density wise Samsung's 4nm is closer to TSMC's 6nm.

 

[mgoldsmith1979]

So how do you differentiate 8+ Gen 1 from Mediatek’s version? ARM cores, same node, same performance, same foundry.

There are still many knobs available for differentiation even with these commonalities. Mediatek has historically skimped on the per-core cache sizing of their ARM implementations, but clearly the different GPU and modem architectures are where the marketing will play out, both in power efficiency and end-device capabilities. But there are reasons why there are so few other market players in this space; Unisoc being the only other large player I'm aware of (excl Exynos and Apple), and they tend to be the value player specific to regional markets more than high-performance. They are where MTK was a decade ago, more or less.

 

[Fred Chen]

TechInsights found 4nm to be actually 5nm for Dimensity 9000 as well: https://www.techinsights.com/blog/mpr-editorial-nanometer-nonsense

 

[Paul2]

Nuvia delayed their desktop chips by 1 year, signals tells they decided to wait for a newer node to overwait the post-COVID blues

 

[IanD]

TechInsights found 4nm to be actually 5nm for Dimensity 9000 as well: https://www.techinsights.com/blog/mpr-editorial-nanometer-nonsense

As usual TechInsights have got the wrong end of the stick...

TSMC N4 is a very small linear optical shrink of N5 (same layouts, same design rules) with some further digital cell library shrinks to reduce area and save a bit of power because of this (higher density = shorter interconnect).

N4P also has some small performance increases (technology tweaks) giving either a bit more speed or lower power, depending on voltage.

It's not a "true 4nm process", it's just a better (smaller, faster, cooler) 5nm-class one -- but TSMC have to call it something, and N4.8 (or whatever) was never going to happen...

 

[Xebec]

It looks like Samsung is planning on showcasing a 3nm processor to President Biden during his visit to South Korea.


However, the rumor is that Samsung's 3nm chip currently only yields 10-20% range!

Source(s): https://www.technoscience.fikrikadim.com/2022/05/20/samsungs-3nm-processor-on-display/

Yield on Samsung's 3nm chip production reportedly drops to a 10-20% range

A report released today says that Samsung's yield on the production of its 3nm Gate-all-around chipsets is a poor 10% to 20%, even worse than the 35% yield on its 4nm chips.

www.phonearena.com

Have to ask - is this a trusted source for this forum?

(I don’t doubt that Samsung yields are low as they appear to have a culture of shouting ‘first!’ at all costs, but .. )

 

[Fred Chen]

As usual TechInsights have got the wrong end of the stick...

TSMC N4 is a very small linear optical shrink of N5 (same layouts, same design rules) with some further digital cell library shrinks to reduce area and save a bit of power because of this (higher density = shorter interconnect).

N4P also has some small performance increases (technology tweaks) giving either a bit more speed or lower power, depending on voltage.

It's not a "true 4nm process", it's just a better (smaller, faster, cooler) 5nm-class one -- but TSMC have to call it something, and N4.8 (or whatever) was never going to happen...

Maybe something like N5S or N5+ would be less misleading (at least to TechInsights).

 

[IanD]

Maybe something like N5S or N5+ would be less misleading (at least to TechInsights).

IIRC TSMC do refer to N4P as "the fourth generation of our 5nm process family" in technology presentations, it really *is* just a marketing name

 

[nghanayem]

IIRC TSMC do refer to N4P as "the fourth generation of our 5nm process family" in technology presentations, it really *is* just a marketing name

All of these node names are

 

[Daniel Nenni]

All of these node names are

IIRC TSMC do refer to N4P as "the fourth generation of our 5nm process family" in technology presentations, it really *is* just a marketing name

True but you have to call them something and as long as the other foundries are on the same page it seems fair enough. CMOS was much easier since you could accurately measure. FinFETs not so much and GAA will also be difficult to measure.