MY NM IS BETTER THAN YOUR NM
On paper, it seems as though TSMC is way ahead of Intel. After all, it’s been over two years since the first commercially available devices powered by 5nm TSMC chips hit the market. Meanwhile, TSMC is undwerway with mass production of 3nm chips. And Intel? Its most advanced node for commercial chips is 10nm. Oh dear!
But hold on. Making assumptions based on the nomenclature involved is problematic. Put another way, TSMC’s nanometers are not directly comparable to Intel’s. How can that be possible? The name attached to a given process node has become detached from the physical reality of the chips involved. It used to relate to the smallest feature inside a chip. Now node naming is more branding and marketing than science.
In the end, what matters is transistor density or the number of transistors any process node can cram into a given area of silicon wafer. More transistors mean more features and performance, while smaller transistors use less power. It’s a win-win as feature sizes shrink. Such are the advances in that regard, current process nodes are measured in a barely comprehensible metric of hundreds of millions of transistors per square millimeter. Yes, completely crazy.
Anyway, when you compare transistor density, it quickly becomes clear that TSMC’s advantage over Intel isn’t quite as overwhelming as that 10nm versus 5nm or even 3nm comparison implies. Intel’s 10nm node delivers just over 100 million transistors per square millimeter, while TSMC’s 10nm clocks in at 53 million. Indeed, TSMC’s 7nm node is only good for 97 million transistors per millimeter squared. Suddenly, Intel’s rebranding of its 10nm node to ‘Intel 7’ seems pretty reasonable.
Of course, TSMC has been making chips on its own 5nm for a couple of years. TSMC 5nm boasts a density of 173 million transistors, well ahead of Intel 10nm.
Meanwhile, TSMC 3nm will up that to a hefty 290 million. Even if Intel manages to get its new 7nm node, branded Intel 4, going next year, it will only edge ahead of TSMC 5nm with 180 million transistors per millimeter squared, placing Intel three years behind TSMC for transistor density.
As for Samsung, its 10nm node was comparable with TSMC for density. But since then it has fallen behind. And what about the new 3nm node that Samsung proudly announced was already in mass production earlier this summer? It’s actually slightly less dense than TSMC 5nm.
As for whether Intel can catch up, well, that’s the million dollar – or should that be transistor – question. Intel’s 5nm tech, which will be marketed as Intel 20A where the ‘A’ stands for angstroms, isn’t due until at least 2024, putting it 18 months behind TSMC at minimum. Intel’s 3nm node, the branding for which hasn’t been announced, is expected to deliver a staggering 520 million transistors per millimeter squared, slightly ahead of TSMC’s 2nm which is forecast to hit 490 million.
If TSMC and Intel both hit their targets, we can expect 2nm TSMC chips in 2025 and Intel closing the gap to just one year with its 3nm process in 2026. And that’s a best-case scenario for Intel. It’s a long road back, that’s for sure.