Hacker News Reader

Comment by tux1968 | Hacker News Reader

tux1968

This seems like a more gentle way to say that innovation in the sector is slowing a bit. Which would account for why "mature" nodes will be producing more revenue than the latest developments coming online.

View on HN

28 nm was a turning point in cost per transistor. Previously, unit cost of transistors declined with gate lengths, but smaller finfet nodes and EUV add enough complexity that cost per transistor is actually rising on more advanced nodes.

The good thing about this is that capacity in "cutting-edge" fans will be lower and lower every year as more products find a "home" node that balances price and performance.

I'm not sure where that idea came from. The last we saw was that N5 was slightly more expensive than N7, but at the time that info came out nobody but Apple could run N5 except for super expensive proto runs that even then were under tight competition. And even then, N5 is still cheaper than N10. Everything else is cheaper for gate count as you go down in gate size, and we should expect N5 to drop in price to under N7.

https://www.tomshardware.com/news/tsmcs-wafer-prices-reveale...

Maybe this idea came from the industry not fully internalizing the consolidation in the space and how much competition for wafers affects the initial price of a new node?

These processes are explicitly not about logic parts like CPUs or GPUs, but analog/power/mixed-signal parts, which don't benefit from smaller node sizes. Many of these parts are made on 0.12/0.18 µm or even bigger nodes because anything smaller is pointless.

> Generally, TSMC’s speciality technologies portfolio includes MEMS, CMOS Image Sensor, Embedded NVM, RF, Analog, High Voltage, and BCD-Power processes.

"Innovation" in the sector always came at tremendous cost. The startup cost for the foundry for a new node has been increasing every generation (requires more wafers and time to amortize over), and the cost per wafer (on a production basis, not including the amortized development cost) has been increasing over time as well (and you're going to be competing with AMD, Nvidia, and Apple for fab space... those are deep pockets to fight against). And finally, the (per customer) development cost at each node is becoming more expensive as well.

All of this means theres tremendous value in exploiting mature processes. And it's going to continue to be this way for quite a long time based on all the trendlines and projections that the big players put out.

It's not just the nodes AFAIK. There's huge demand for EUV machines but they are lower margin products for the company that makes them.

I don't really understand this business, or any business really, so don't take my word for it

I have no digits, but it seems that at least bruteforcing the volumes is in ways with how much chips Apple + AMD + Qualcomm ship.

EUV productivity is a big problem. Currently machines are stopped every few shifts for cleaning tin from mirrors. Even without that, light sources never managed to live to the original power output promise.

There are few known ways to produce EUV light besides tin plasma.

One is in using some plasmonic quantum processes, I can't even fathom how to describe, to produce microwatts of EUV. The other is a free electron laser to produce potentially kilowatts of it non-stop.

The later approach requires to build the whole fab around a synchrotron, a commitment currently nobody can take even with current surreal capex numbers.

Maybe we should rebuild the Large Hadron Collider as the core of a silicon fab/s

Particle pipes for semi fabs (a la oil pipelines) would be a cool bit of retrofuturist-y world building

Oh they could do the latter, they just need to trust it's gonna work.

There is such a facility, not far away from me. Though not that much free space around it anymore.

https://www.xfel.eu/

https://openstreetmap.de/karte.html?zoom=17&lat=53.58825&lon...

Equipment for ion implantation already includes mini accelerators [1] [2]. The semiconductor equipment industry in general has many machines that feel like they came out of a physics lab into a semi fab. For example, plasma dry-etching or deep-reactive ion etching. EUV litho is just one of many very interesting problems--currently the bottleneck so it's talked about a lot.

[1] https://en.wikipedia.org/wiki/Ion_implantation [2] https://www.axcelis.com/products/purion-xe-series-high-energ...

edit: add link to cool ion implanter machine pics

> There's huge demand for EUV machines but they are lower margin products for the company that makes them.

ASML, a Dutch company, is at the core of TSMCs success. But the machines are only one part of the ecosystem... the buildings to house them are incredibly expensive because you need the cleanest of clean rooms you can get, and then you need expert staff to operate these, and since there isn't much of a domestic chip industry any more in the Western countries, that knowledge has long ago moved off to Taiwan and South Korea.

And on top of that, Western wages are extremely high and environmental regulations are extremely strict after half of Silicon Valley became a superfund site from all the stuff that leeched into the soil from decades of chipmaking.

You also need hundreds of other machines from Applied Materials, KLA Tencor, Lam Research, Tokyo Electron, etc. Then, years of process development R&D: what temperature do we bake this layer at, how long, what profile, what atmosphere in the tool… Every process step has a large parameter set to optimize. Creating a process is a painstaking many many variable optimization slog.

Just wanted to add that since I see people only mention ASML often here. They are very important, but there is so much more to TSMC’s success.

I disagree. A lot of the upcoming requirements of transistors just do not need small transistor nodes. For example, for RF applications, power handling needs simply require larger gate lengths. There is no way you can use a 7nm finfet to realize some RF applications.

Innovation in semiconductors takes many forms, not just pushing to lower transistor nodes.

Reply on news.ycombinator.com