Not Just Numbers: How TSMC’s “Small-Step, Fast-Run” Strategy Became Its Secret Weapon Against Samsung and Intel

[karin623]

Everyone knows TSMC as the world’s most advanced chipmaker. But few realize that one of its biggest advantages isn’t measured in nanometers — it’s in how those nanometers evolve.

In a rare semi-official interview with Y.J. Mii, TSMC’s newly appointed Executive Vice President and Co-Chief Operating Officer, we finally get a glimpse into the company’s long-kept secret: the half-node.

Born in the 1990s, this deceptively simple idea — making small, incremental shrinks instead of big leaps — became the cornerstone of TSMC’s “small-step, fast-run” strategy. It’s how the company turned caution into precision, iteration into dominance, and consistency into innovation.

From the legendary 28 nm era that Morris Chang called his proudest victory, to the current 4 nm and 5 nm processes powering AI data centers, the same philosophy endures. The half-node isn’t just a tweak in chip design — it’s the rhythm that keeps TSMC perpetually ahead of Samsung and Intel.

Read the full story: Inside TSMC’s Secret Weapon: The Power of the Half-Node.

Not Just Numbers: How TSMC’s “Small-Step, Fast-Run” Strategy Became Its Secret Weapon Against Samsung and Intel

Liang-rong Chen

cwnewsroom.substack.com

 

[hist78]

Everyone knows TSMC as the world’s most advanced chipmaker. But few realize that one of its biggest advantages isn’t measured in nanometers — it’s in how those nanometers evolve.

In a rare semi-official interview with Y.J. Mii, TSMC’s newly appointed Executive Vice President and Co-Chief Operating Officer, we finally get a glimpse into the company’s long-kept secret: the half-node.

Born in the 1990s, this deceptively simple idea — making small, incremental shrinks instead of big leaps — became the cornerstone of TSMC’s “small-step, fast-run” strategy. It’s how the company turned caution into precision, iteration into dominance, and consistency into innovation.

From the legendary 28 nm era that Morris Chang called his proudest victory, to the current 4 nm and 5 nm processes powering AI data centers, the same philosophy endures. The half-node isn’t just a tweak in chip design — it’s the rhythm that keeps TSMC perpetually ahead of Samsung and Intel.

Read the full story: Inside TSMC’s Secret Weapon: The Power of the Half-Node.

Not Just Numbers: How TSMC’s “Small-Step, Fast-Run” Strategy Became Its Secret Weapon Against Samsung and Intel

Liang-rong Chen

cwnewsroom.substack.com

The half-node and small-step approach improves on time product delivery. Predictability in product delivery is extremely important to TSMC and its customers, such as Apple, Qualcomm, MediaTek, Nvidia, AMD, and the broader supply chain and ecosystem.

Apple relies on predictability to plan its finances for manufacturing, marketing, and business expansion. Foxconn needs it to hire and train employees and to prepare the necessary, often expensive, production tools. AMD depends on it to inform its customers when they can roll out new data center servers to replace Intel based systems.

 

[benb]

TSMC guessed right on 28nm (gate last, HKMG) and iterated steadily from there. They also got Apple from 20nm onward. Pulling Apple away from Samsung destroyed Samsung (in hindsight).

Apple puts the wind in TSMC sails.

 

[Daniel Nenni]

TSMC guessed right on 28nm (gate last, HKMG) and iterated steadily from there. They also got Apple from 20nm onward. Pulling Apple away from Samsung destroyed Samsung (in hindsight).

Apple puts the wind in TSMC sails.

We have been doing half node for many years. High volume fabless companies demanded it. Now we call it a family of processes but they really are half nodes.

I remember doing an optical shrink between 65nm and 40nm with ATI GPUs. It was an optical shrink with some extra yield enhancements IIRC.

 

[Xebec]

We have been doing half node for many years. High volume fabless companies demanded it. Now we call it a family of processes but they really are half nodes.

I remember doing an optical shrink between 65nm and 40nm with ATI GPUs. It was an optical shrink with some extra yield enhancements IIRC.

I was going to ask how long this has been prevelant. I remember following the "standard" node naming conventions of Intel, and AMD.. then ITRC.

Then I started watching GPUs. When Intel was going from 65nm to 45nm, TSMC also had a 55nm step in there (65 --> 55 --> 40). I also vaguely remember TSMC also having a 150nm, straddling Intel's 180nm and 130nm nodes. (.18, .13 micron ).

Were these early half nodes customer requests? or just smart business based upon the equipment and cost at the time?

 

[tomatoma]

Half nodes were established due to various cost-related reasons.
This is not just a TSMC issue.

 

[tomatoma]

The half-node and small-step approach improves on time product delivery. Predictability in product delivery is extremely important to TSMC and its customers, such as Apple, Qualcomm, MediaTek, Nvidia, AMD, and the broader supply chain and ecosystem.

Apple relies on predictability to plan its finances for manufacturing, marketing, and business expansion. Foxconn needs it to hire and train employees and to prepare the necessary, often expensive, production tools. AMD depends on it to inform its customers when they can roll out new data center servers to replace Intel based systems.

I don't really understand

 

[hist78]

I don't really understand

About what?

 

[Artificer60]

TSMC guessed right on 28nm (gate last, HKMG) and iterated steadily from there.

??? Why guess? Intel put High-K metal gate into production in 2007 and the tear downs made it clear right away it was gate last. I distinctly remember seeing reverse engineering reports that identified this. TSMC didn't introduce High-K metal gates until 2009. I'm sure TSMC had contingency plans for both gate first and gate last in development. Once Intel landed it, I believe TSMC was pragmatic enough to choose the obvious path.

 

[BruceA]

??? Why guess? Intel put High-K metal gate into production in 2007 and the tear downs made it clear right away it was gate last. I distinctly remember seeing reverse engineering reports that identified this. TSMC didn't introduce High-K metal gates until 2009. I'm sure TSMC had contingency plans for both gate first and gate last in development. Once Intel landed it, I believe TSMC was pragmatic enough to choose the obvious path.

Glad you cleared the air, intel was first for every innovation. If Intel was to go away the world will be left with a bunch of half nodes and soon quarter nodes. I am not at all being negative, but TSMC is yet to show they can do real innovation. They are great at building the ecosystem, ramping at scales and doing the crazy hard work that simply doesn't exist in western culture. The deep culture of both the company and the Taiwan culture set the amazing time and place to enable the perfect match. The things that the engineers and culture squeeze out of their tools and the precision that they force are beyond the imagination of anyone outside of the company and even their tool vendors.

Intel was first to strained silicon, first to the right way to to High-K, first to finfet, first to GAA. One could argue that their ego and desire for innovation sometimes blinds them to the real wants and business practicality of what their customers want. The challenge for them is they have no corner on innovation and TSMC is simply relentless and has a workforce and culture of nose to the grindstone for manufacturing and RD that doesn't exist in team Blue. It will be hard for LBT to change as he simply has no clue on the fab side of things.

After losing leadership at 10nm catching up and losing that scale advantage and not having any real customers besides their own is a huge boat anchor. Let's hope LBT can fix that, so far the jury is out on the foundry side for that and as I sad LBT has no clue to the fab side.

 

[hist78]

first to GAA.

Toshiba was the first to demonstrate a Gate-All-Around (GAA) transistor in 1988. And Samsung was the first to achieve volume production of 3nm products with GAA in June 2022., although the yield was bad.

IMO, both Intel and TSMC have made tremendous innovations. Some of TSMC’s breakthroughs, such as immersion lithography and advanced packaging, have become the very foundation of today’s semiconductor industry.

TSMC’s innovative pure-play foundry business model disrupted the entire industry and helped turn the dreams of fabless companies into reality. Those big and small fabless firms (or divisions) have since developed countless great products and technologies, gradually eroding the competitiveness of traditional IDMs.

 

[siliconbruh999]

Intel is also first to GAA+BSPN in HVM as for advanced packing Intel has also made breakthrough what about Glass Substrates and Silicon Photonics Intel has tons of R&D in those stuff but nowadays TSMC gets publicity like they are the chosen one.

 

[Daniel Nenni]

Toshiba was the first to demonstrate a Gate-All-Around (GAA) transistor in 1988. And Samsung was the first to achieve volume production of 3nm products with GAA in June 2022., although the yield was bad.

IMO, both Intel and TSMC have made tremendous innovations. Some of TSMC’s breakthroughs, such as immersion lithography and advanced packaging, have become the very foundation of today’s semiconductor industry.

TSMC’s innovative pure-play foundry business model disrupted the entire industry and helped turn the dreams of fabless companies into reality. Those big and small fabless firms (or divisions) have since developed countless great products and technologies, gradually eroding the competitiveness of traditional IDMs.

Samsung 3nm had no customers due to a horrible PDK so 3nm was later renamed to Samsung 2nm. In my opinion Intel was first to HVM with GAA and TSMC will be a close second. I only count a process node when products ship. I do not count Intel 20A for example so there was no 5N4Y.

TSMC did develop 28nm HKMG in both gate-first and gate-last then chose the one that yielded best which was gate-last. TSMC also did 28nm PolySiON. Parallel R&D is common for TSMC. Intel was certainly first to HKMG but TSMC also did the work. IBM, UMC, Chartered, and Samsung (Common Platform) used the IBM 28nm recipe (Gate-first HKMG) and did not yield which is why TSMC went on 28nm allocation. TSMC planned for a 40% 28nm market share and ended up with twice that.

Now that TSMC has no real competition capacity planning amongst customers is not as hard.

 

[siliconbruh999]

Intel did 4 CEO in 5 Months

 

[Artificer60]

TSMC did develop 28nm HKMG in both gate-first and gate-last then chose the one that yielded best which was gate-last. TSMC also did 28nm PolySiON. Parallel R&D is common for TSMC. Intel was certainly first to HKMG but TSMC also did the work. IBM, UMC, Chartered, and Samsung (Common Platform) used the IBM 28nm recipe (Gate-first HKMG) and did not yield which is why TSMC went on 28nm allocation. TSMC planned for a 40% 28nm market share and ended up with twice that.

It was never my intent to imply that TSMC just rode on Intel's coat tails. The fact that they have moved from being a fast follower to a leadership role shows that they are more than capable of innovation. Frankly, I think a viable Intel foundry effort + TSMC would lead to some amazing advances.

 

[Artificer60]

Glad you cleared the air, intel was first for every innovation. If Intel was to go away the world will be left with a bunch of half nodes and soon quarter nodes. I am not at all being negative, but TSMC is yet to show they can do real innovation.

Not sure how you define real innovation, but I would argue that since the end of Dennard scaling, every node has required significant innovation.

After losing leadership at 10nm catching up and losing that scale advantage and not having any real customers besides their own is a huge boat anchor. Let's hope LBT can fix that, so far the jury is out on the foundry side for that and as I sad LBT has no clue to the fab side.

I would quote Andrew Carnegie here. "I owe whatever success I have attained to my ability to surround myself with people who are smarter than I am" I believe Lip-Bu Tan has his ego sufficiently in check that he is capable of doing that and will look at the results and make changes as needed.

 

[samwilde]

I do not count Intel 20A for example so there was no 5N4Y.

Even if you don't, a node every single year is still impressive. Especially considering how Intel did before Gelsinger came back.

I think they did well in skipping 20A when they figured out 18A was yielding. Had it proved difficult to develop 18A, 20A would have been the backup.

Switching to market available tools, instead of internal ones, and regular PDKs was also long overdue. Even if the transition has proven to be difficult.

 

[siliconbruh999]

The problem with 20A was money and the majority of lineup was N3B anyway so it ended up being cut it didn't make logical or financial sense at that point of time.