TSMC Update: 2nm in Development, 3nm and 4nm on Track for 2022
by Anton Shilov on April 26, 2021 2:15 PM ESTFor TSMC, being the world's largest foundry with nearly 500 customers has its peculiarities. On the one hand, the company can serve almost any client with almost any requirements. On the other hand, it has to stay ahead of everyone else both in terms of capacity and in terms of technology. As far as capacity is concerned, TSMC is unchallenged and is not going to be for years to come. As for fabrication technologies, TSMC has recently reiterated that it's confident that its N2, N3, and N4 processes will be available on time and will be more advanced than competing nodes.
Confidence
Early this year TSMC significantly boosted its 2021 CapEx budget to a $25 – $28 billion range, further increasing it to around $30 billion as a part of its three-year plan to spend $100 billion on manufacturing capacities and R&D.
About 80% of TSMC's $30 billion capital budget this year will be spent on expanding capacities for advanced technologies, such as 3nm, 4nm/5nm, and 6nm/7nm. Analysts from China Renaissance Securities believe that most of the money on advanced nodes will be used to expand TSMC's N5 capacity to 110,000 ~ 120,000 wafer starts per month (WSPM) by the end of the year. Meanwhile, TSMC said that 10% of its CapEx will be allocated for advanced packaging and mask making, whereas another 10% will be spent on specialty technologies (which includes tailored versions of mature nodes).
TMSC's the most recent CapEx hikes announcements were made after Intel announced its IDM 2.0 strategy (that involves in-house production, outsourcing, and foundry operations) and to a large degree reaffirms TMSC's confidence in both short-term and long-term future even ahead of intensified competition.
"As a leading pure-play foundry, TSMC has never been short on competition in our 30-plus-year history, yet we know how to compete," said C.C. Wei, president and CEO of TSMC, at a recent conference call with analysts and investors. "We will continue to focus on delivering technology leadership, manufacturing excellence, and earning our customers' trust. The last point, customers' trust, is fairly important because we do not have internal products that compete with customer."
Advertised PPA Improvements of New Process Technologies Data announced during conference calls, events, press briefings and press releases |
|||||||||
TSMC | |||||||||
N7 vs 16FF+ |
N7 vs N10 |
N7P vs N7 |
N7+ vs N7 |
N5 vs N7 |
N5P vs N5 |
N4 vs N5 |
N3 vs N5 |
||
Power | -60% | <-40% | -10% | -15% | -30% | -10% | lower | -25-30% | |
Performance | +30% | ? | +7% | +10% | +15% | +5% | higher | +10-15% | |
Logic Area Reduction % (Density) |
70% |
>37% |
- |
~17% |
0.55x -45% (1.8x) |
- |
? | 0.58x -42% (1.7x) |
|
Volume Manufacturing |
2018 | 2018 |
2019 | Q2 2019 |
Q2 2020 | 2021 | 2022 | H2 2022 |
N5 Gaining Customers
TSMC was the first company to start high volume manufacturing (HVM) of chips using its N5 (5 nm) process technology in mid-2020.
Initially, the node was used solely for TSMC's alpha customers — Apple and HiSilicon. Shipments to the latter ceased on September 14, which left all of the leading-edge capacity to Apple. By now, more customers are ready with their N5 designs, so the adoption of this node is growing. Meanwhile, TSMC says more customers are planning to use N5 family of technologies (including N5, N5P, and N4) than it expected just several months ago.
"N5 is already in its second year of volume production with yield better than our original plan," said Mr. Wei. N5 demand continues to be strong, driven by smartphone and HPC applications, and we expect N5 to contribute around 20% of our wafer revenue in 2021. […] In fact, we are seeing stronger engagement with more customers on 5 nm and 3 nm [versus 7 nm at similar stages]. The engagement is so strong that we have to really prepare the capacity for it."
For TSMC, HPC applications include many different types of products, including AI accelerators, CPUs, GPUs, FPGAs, NPUs, and video gaming SoCs, just to name a few. Since they're just a contract manufacturer, TSMC does not disclose what kinds of products it makes using one node or another (we do know that it builds the Apple A14 SoC for smartphones/tablets/STBs as well as the Apple M1 SoC for PCs and tablets), but the very fact that adoption of N5 is growing in the HPC segment is important.
"We expect demand for our N5 family to continue to grow in the next several years, driven by the robust demand for smartphone and HPC applications," the head of TSMC said. "We expect to see HPC, not only in the first wave, but in additional waves of demand to support our leading [N5] node in the future, actually."
It is not particularly surprising that TSMC's N5 is gaining market share among adopters of leading-edge technologies. Analysts from China Renaissance estimate that TSMC's N5 features a transistor density of around 170 million transistors per square millimeter (MTr/mm2), which if accurate, makes it the densest technology available today. By contrast, Samsung's Foundry's 5LPE can boast with about 125 MTr/mm2 ~130 MTr/mm2, whereas Intel's 10 nm features an approximately 100 MTr/mm2 density.
In the coming weeks TSMC is set to start making chips using a performance-enhanced version of its N5 technology called N5P that promises to increase frequencies by up to 5% or reduce power consumption by up to 10% (at the same complexity). The technology offers a seamless migration path for customers without requiring significant engineering resource investment or longer design cycle time, so anyone with an N5 design can use N5P instead. For example, early adopters of N5 could re-use their IP for their N5P chips.
N4: On Track for Next Year
TSMC's N5 family of technologies also includes evolutionary N4 process that will enter risk production later this year and will be used for mass production in 2022.
This technology is set to provide further PPA (power, performance, area) advantages over N5, but keep the same design rules, design infrastructure, SPICE simulation programs, and IPs. Meanwhile, since N4 further extends usage of EUV lithography tools, it also reduces mask counts, process steps, risks, and costs.
"N4 will leverage the strong foundation of N5 to further extend our 5 nm family," said Mr. Wei. "N4 is a straightforward migration from N5 with compatible design rules while providing further performance, power and density enhancement for the next wave of 5-nanometer products. N4 risk production is targeted for second half this year and volume production in 2022."
By the time N4 enters HVM in 2022, TSMC will have about two years of experience with N5 and three years of experience with EUV. So expectations are that yields will be high and the performance variability promises to be low.
But even as cutting-edge as N4 is slated to be, it's not going to be the most advanced fabrication technology that TSMC will offer next year.
N3: Due in H2 2022
In 2022, the world's largest contract maker of chips will roll out its brand-new N3 manufacturing process, which will keep using FinFET transistors, but is expected to offer the whole package of PPA improvements.
In particular, versus their current N5 process, TSMC's N3 promises to increase performance by 10% – 15% (at the same power and complexity) or reduce power consumption by 25% – 30% (at the same performance and complexity). All the while the new node will also improve transistor density by 1.1 ~ 1.7 times depending on the structures (1.1X for analog, 1.2X for SRAM, 1.7X for logic).
N3 will further increase the number of EUV layers, but will keep using DUV lithography. Also, since the technology keeps using FinFET, it will not require a new generation of electronic design automation (EDA) tools redesigned from scratch and development of all-new IPs, which might become a competitive advantage over Samsung Foundry's GAAFET/MBCFET-based 3GAE.
"N3 will be another full node stride from our N5 and will use FinFET transistor structure to deliver the best technology maturity, performance, and cost for our customers," said Mr. Wei. "Our N3 technology development is on track with good progress. We continue to see a much higher level of customer engagement for both HPC and smartphone applications at N3 as compared with N5 and N7."
In fact, TSMC's claims about growing customer engagement with N3 indirectly telegraphs its high expectations for N3.
"[N3] risk production is scheduled in 2021," said TSMC's CEO. "The volume production is targeted in second half of 2022. Our N3 technology will be the most advanced foundry technology in both PPA and transistor technology, when it is introduced. […] We are confident that both our [N5] and [N3] will be large and long-lasting nodes for TSMC."
Beyond N3
Gate-all-around FETs (GAAFETs) are still a part of TSMC's development roadmap. The company is expected to use a new kind of transistors with its 'post-N3' technology (presumably N2). In fact, the company is in path-finding mode for next generations of materials and transistor structures that will be used many years down the road.
"For advanced CMOS logic, TSMC’s 3nm and 2nm CMOS nodes are progressing nicely through the pipeline," the company said in its annual report recently. "In addition, TSMC's reinforced exploratory R&D work is focused on beyond-2nm node and on areas such as 3D transistors, new memory and low-R interconnect, which are on track to establish a solid foundation to feed into many technology platforms.
It is noteworthy that TSMC is expanding capacity for R&D operations at Fab 12, where N3, N2, and more advanced nodes are currently being researched and developed.
Summary
Overall, TSMC is confident that its "everyone's foundry" strategy will enable it grow further in terms of scale, market share, and sales. The company also expects to maintain its technology leadership going forward, which is pivotal for growth.
"For the full year of 2021, we now forecast […] foundry industry growth [at] about 16%," said Wendell Huang, CFO of TSMC, at a recent conference call with analysts and investors. "For TSMC, we are confident we can outperform the foundry revenue growth and grow by around 20% in 2021."
The company has a strong technology roadmap and it is set to continue introducing improved leading-edge nodes every year, thus offering its customers improvements at a predictable cadence.
TSMC knows how to compete against rivals with leading-edge nodes as well as makers of chips focused on specialty process technologies, so it does not see Intel Foundry Services (IFS) as an immediate threat especially because the blue giant is going primarily after leading-edge and advanced nodes.
Financial analysts generally share TSMC's optimism mainly because of the expectation that the company's N3 and N5 nodes are not going to have competitors offering similar transistor densities and wafer starts.
"Following Intel's announced foundry comeback in March, TSMC’s willingness to set a 3-year $100 billion CapEx/R&D investment plan, starting from 2021, indicates its confidence to widen its foundry leadership," Szeho Ng, an analyst with China Renaissance Securities. "We see TSMC’s strategic value rising with N3/N5: strong N5 tape-out activities from HPC/smartphone applications and more N3 client engagement vs N5/N7 at similar stages."
74 Comments
View All Comments
melgross - Tuesday, April 27, 2021 - link
Yeah, sure. Just like s86 couldn’t do any useful work in the workstation days. What happened there? Just like laptops could never do any useful work. There too. Now we see Apple’s M1, the worst chip Apple will ever produce for desktops and high end tablets, run rings around most x86 laptops and even many desktops. It will just get worse for s86 as manufacturers continue to look at that and figure that they need to produce better chips too. Qualcomm isn’t sitting still either.Keep living in your made up world.
Silver5urfer - Tuesday, April 27, 2021 - link
Ah melgross, I remember you on the same context last time. Now coming to your points M1 gets crushed by Ryzen processors with SMT granted the only advantage is thin and light BS but when the workload increases that M1 cannot compete."run rings around most x86 laptops and even many desktops" many desktops lol what a massive pile of bs. Which desktop processor got beaten by M1 ? Can you point out the benchmarks and realworld test cases, will you say it beat 7700K in the real world performance or any 4790K processor in gaming or such, it doesn't even have a damn GPU lol. A stupid iGPU cannot do anything vs a PCIe PEG GPU.
M1 has dedicated blocks for encode and decode which is why they got good performance in Encoding vs x86 processors since they do not have that functionality inbuilt, but that's nothing if NVENC comes into play, it's game over. And if we add the Matlab and other compute workloads for consumer the HT/SMT based x86 processors will melt the M1 garbage which can be only fit for Apple first party workloads which were made by Adobe, MS and other companies and their Final Cut Pro and running their joke of an iOS apps on that with jank (see LTT new video on his iOS apps running on that M1)
Qualcomm is not going to do anything, they abandoned the idea of Server since Centriq was axed and their custom uArch team got axed with that, that SD888 still fall behind AT's great SPEC graphs but in realworld scenarios it doesn't even have any impact. There's absolutely none, Qcomm 888 processor phones vs iPhone have bog standard same response times in opening and using applications, which is called "realworld" workloads.
"x86 couldn't do any useful work in Workstation days" HEDT or the Enterprise workloads ? Is that a joke lol.
Keep shilling for Apple when having no absolute advantage of them in the consumer space. Esp the fact that Mac makes only a pathetic 10% of Apple revenue AND the OS marketshare at 13%.
And the fact that ARM processors are simply custom and Enterprise custom centric which do not have any use case for end users, what use you have with ARM processor when it's not even there to replace the x86 desktop processor, as for the IT, EPYC Milan 7763 is the king of the hill which even a normal person can buy and build a homelab, speaking of home server setups, which processor is to buy ? M1 ? haha.
And Xilinx / FPGAs are not going to sit idle, that's going to be the biggest bet for AMD, Intel already has FPGAs with them, that compute block will be having any customized workload which is where many AI and other things will be added to the x86 CPUs.
I don't really understand why people shill for Apple and claim some magical bs, before M1 there's nothing for consumer ARM yet saw so many posts on that BS claims of x86 is dead while using it daily. Peak comedy.
Matthias B V - Tuesday, April 27, 2021 - link
What does the GPU have to do with it - It is not like you can't design a ARM CPU that connects to GPU over PCIe... Nvidia will do so and Imagination Technologies might also try to get into that space.Even AMD might be ok with RDNA/CDNA + ARM combination. They will milk x86 but have less issues moving to ARM compared to Intel.
Intel is the one that has most to lost on ARM / RISC as they hold the x86 keys and dominate market.
Matthias B V - Tuesday, April 27, 2021 - link
Sorry but x86 is beyong prime. When I wrote dead I dead not mean it is not existing but by that time it lost significant market share and will continue losing basically marking its long term end / decline. Why? Well:Server / HPC: We can already see multiple companies pushing into the market and with ARM v9 a quite well hardware base is provided. From there on when Software support increases it is set. How many companies already started own designs!
Mobile: How many users even today only use a tablet and no real Notebook anymore. How many popular games [especially in Asia] do not need massive GPUs / CPUs. Perfromance will be enough for both and they prefer battery lifetime and ARM. Once Microsoft stops screwing it up it will also increase share fast.
Desktop: Yes share is low and will be probably the longest time a x86 domination. But the market is nothing compared to mobile and server.
And how many people will switch to newest MacBook Airs and iPadPros taking lots of x86 away as they just like the battery life and performance of newest M1 devices combined with easy to use OS - It is just enough for most everyday users and prosumers besides gamers!
x86 is everything but lean and smart if you compare SSE, AVX, AV2 AVX512 compared to SVE... K12 might have not worked because time wasn't ready for it yet. But now it is and also AMD was starving cash and couldn't push it. It is not the same situation as is now.
Matthias B V - Tuesday, April 27, 2021 - link
Also one thing about market share: We talk about desktop, server, mobile.With mobile we talk notebooks and tablets however smartphones are forgotten!
If include those and check for CPU / APU market share then you can already see a quite big possible share lost...And as I said - This perfromance of those tablets and phones is ebough for most people that are not gaming or professional oriented!
Silver5urfer - Tuesday, April 27, 2021 - link
Lmao, Server and Mobile are you joking ? Server space is x86 dominated over 90% of Intel and AMD Is clawing at that with ARM restricted to AWS ONLY and there are no others offering the ARM HW at the moment there will be in the future yes as of today's New ARM Neoverse news. But once ARM has a solid position AND beating out x86 then we will talk again.Oh mobiles !! Amazing pieces of trash HW outselling Desktops. What a revelation, dude we talk on x86 performance and the capabilities of what they can do, your stupid iPhone or my Android phone which cannot do anything remotely compared to what x86 notebooks and desktops can do. You don't even have a filesystem to control forget running applications and scientific or machine learning workloads or physics either. Basically browsing Internet and using Social Media crap and you are using that to define the ARM dominance ? what a fcking joke.
"How many people will switch" lol, Mac has 10% marketshare and 10% world wide OS netmarketshare, what are you babbling about the world flipping to Apple HW and SW in a fortnight or magical moment. And iPad cannot translate x86 code it doesn't and cannot since it's not running Mac OS period. Only M1 Mac can do that. Everything is in the future your talk is not just dumb but really a time waste, idk from where all these Apple shills come from and claim the x86 is dead, damn it.
What a load of bs is this "It is not like you can't design a ARM CPU that connects to GPU over PCIe... Nvidia will do so and Imagination Technologies might also try to get into that space.
Even AMD might be ok with RDNA/CDNA + ARM combination. They will milk x86 but have less issues moving to ARM compared to Intel."
Sigh, what is that ? Imagination IP is stolen by Apple and they are dead sold out to a Chinese company and so far eGPU doesn't exist and why the fuck is Apple only one ? What about others, yeah you don't have anything else, and before M1 also same argument but nothing, now also same. And there's no product that exists on the market that is offering ARM CPUs to be paired with the PCIe express slots, fucking there's no PCIe ecosystem for ARM Consumer CPUs which is M1 soldered POS how are you even thinking about a GPU lol, and AMD is moving to ARM ? WTH. Last time Lisa Su said they do not have any plans for BigLittle forget ARM, there's one rumor today that came about APUs of Zen 5 using Big Little, that is a rumor and they are competing against Intel x86 big little ADL rather than stupid M1 which has pathetic marketshare.
That SVE is going to be ground shattering moment of truth. Ah I see, maybe it will make PS3 cell emulation go nuts and wreck havoc onto the x86, let us see. What about x86 market Xilinx FPGA ? nah that's dumb since it's x86. And x86 even though it's CISC it uses RISC under just like ARM just the thing is ARM processors have wide architecture on front end where as x86 use high clockspeed and SMT. But let's ignore all that, x86 is junk all the while relying on it, I bet you typed this on an x86 machine.
What do you think about Homelab ? nah it's junk M1 ftw right...better read up on what people can do on their x86 old Xeon hardware. Not even new. While for ARM what do we have ? Pi that's all, only consumer centric device which can run Linux code natively and can do same home server type compute for small workloads. Nah it's all too much logic, let's only talk Apple and M1 and the never ending bs of x86 is dead ARM is the future.
29a - Wednesday, April 28, 2021 - link
Go away douch bag no one is going to read your wall of idiot text.0razor1 - Tuesday, May 4, 2021 - link
I actually found it weirdly entertaining.melgross - Tuesday, April 27, 2021 - link
Legacy is what keeps Microsoft on x86. Losing legacy is why Microsoft is having so many problems moving to ARM.Arsenica - Monday, April 26, 2021 - link
To the layman it may seem so but marketing "nanometers" have really clouded the waters.TSMC´s N5 has a density of 170Mtr/mm^2 while N3 is projected to do around 285Mtr/mm^2 in 2023.
By then Intel's 7nm (P1278) will do around 235Mtr/mm^2.
So TSMC will have a density advantage but not by nearly as much as marketing nanometers make it appear.