Hacker News Reader

Comment by ein0p | Hacker News Reader

They are already at 7nm (low yield, but you gotta start somewhere), and they plan to invest $140B into this. I wouldn’t be so sure about them “not beating Intel” or other such prognostications. 140B _in China_ buys a lot more “trying” than it does in the West.

View on HN

They also deal with a lot less bureaucracy and inefficient planning process (with decade long litigations) than in the west, at least if they have gvt backing.

That being said foundries are a hard business, there is a reason there has been so much concentration and companies throwing the towel.

Node size is only one piece of the puzzle, and not a sufficient one to beat competitors.

They are still using DUV for that. It will likely take at least a decade to develop EUV independently, which is needed for smaller process nodes.

Not with $140B in a country where literally everything is cheaper, and not when they know it can be done (and likely also have at least half the technical documentation). It won’t take a “decade”.

Apparently they're working on a DUV 3nm process, which is a little insane if you think about it. Would certainly be interesting to see it working, if it does work.

Significantly less time if they simply “acquire” the information needed via industrial espionage :)

EUV is the way that won from the R&D by the current leaders.

It might not be the only way to get things to work, though clearly I have no real idea. :)

Everyone keeps beating this 7nm drum but ignored the fact it’s 7nm performing the equiv of a cpu from like 7-10 years ago…

CPUs were already pretty good 7 years ago though. I still run some of them.

If we follow Moore's Law, that is 3.5 doublings, wish is a little more than 11x difference. Means you need at least 10x the amount of hardware to get the same amount of work done

That doesn’t seem to hold actually. My NAS/VM host runs 14nm Skylake and I see zero reason to upgrade. I don’t care that the die size is larger - I’m never going to see it. Power consumption is also pretty reasonable.

> If we follow Moore's Law ...

At least cpu's haven't been adhering to Moore's Law for a few years now.

Tell that to Intel.

Meanwhile TSMC has 2nm at Apple scale and trialling Nvidia's new computational lithography.

So it's not like the West is sitting around doing nothing.

Intel has bought all the high NA EUV ASML produces until the first quarter 2025 by the way. TSMC is not the west, TSMC is TSMC.

Also China has started investing heavily in domestic AI chips. Huawei's Ascend is roughly comparable to the A100 I think, but you shouldn't forget that this has caused a flurry of GPU as a service company to pop up outside of China that serve GPU compute to Chinese customers.

The best this kind of legislation does is make it a bit more expensive for China to access GPU compute. It's worth remembering that a lot of tech was banned from Iran, but that never stopped Iranians from getting any of those components. It just made them twice as expensive.

https://www.trendforce.com/presscenter/news/20231211-11957.h...

TSMC is not the “West”. It could go away within a week once the festivities begin in earnest.

Reply on news.ycombinator.com