TSMC: SoIC | Hacker News Reader

TSMC: SoIC

3dfabric.tsmc.com

135 pointsblopeur41 comments

M1 chip has a bog standard Package-on-Package architecture - there is nothing special about it. The CoWoS and other 3D packaging technologies from TSMC's marketing pages aren't on the M1 chip. I believe there isn't even a silicon interposer due to cost presumably [1]. It is substrate on top of substrate (PoP) to connect the memory and CPU. CPU+GPU are on the same die which is pretty remarkable (Intel's IGPUs have been on the same die as well albeit they're not as powerful as M1s graphics) but that's somewhat unrelated to the packaging.

Here is a good diagram (sans the silicon interposer): https://www.researchgate.net/figure/3D-stacked-DRAM-example-...

Btw, memory has been stacked like that for a decade or more. Used to be stacked + wirebonded. But now, we have through-silicon vias (TSVs). The reason you can stack memory like that are many, one of them being thermals / power density.

[1] https://www.youtube.com/watch?v=t6KUnC-oU5g

Whats, Whys and Hows of TXMS-SoIC : https://3dfabric.tsmc.com/english/dedicatedFoundry/technolog...

Papers :

* https://ieeexplore.ieee.org/document/8811194

* https://ieeexplore.ieee.org/document/8776486

Too bad they chose a name already taken by another very well established IC packaging technology : https://en.wikipedia.org/wiki/Small_outline_integrated_circu...

SOIC is the external package though, not sure it'll be that confusing as many chips are typically packaged in many different packages.

Is there a risk in one manufacturer seemingly getting so far ahead in terms of being able to deliver high-performance processors?

Intel was there. If a company gets far ahead then they start to rest on their laurels and then competitors catch back up. It's cyclical.

I think what hurt intel in the long run was actually their fully integrated stack.

They couldn't really sell fab time to other companies cuz they didn't have a good attempt at fabbing for 3rd parties. So they were only fabricating intel chips, that's not nearly as big of a business as TSMC. So basically, TSMC can keep their last 3+ gens of fabs open and making product, while intel couldn't really reap any benefit from keeping old process nodes open.

AMD made a good call selling off global foundries when it was apparent they couldnt keep up with Intel even.

I think the really rough issue here is what happens if China does try to take Taiwan (and this is not unrealistic as China has been more aggressive towards them recently).

TSMC is now a strategic asset for the USA. We've been attempting to get them to build a fab in the states for exactly this reason.

Intel fabs plenty of stuff at the 22 nm node, including chipsets for their CPUs[1], RF and other analog uses[2], novel RAM technologies, and their Lakefield/Foreveros thing. Given Intel's expanding product portfolio, there's likely additional internal customers for older nodes.

[1]: https://www.tomshardware.com/news/intel-14nm-shortage-h310c,...

[2]: https://ieeexplore.ieee.org/document/9075914

Is that enough volume to cover all of their older , and large, fab capacity ? i'm not so sure.

Intel has a large number of fabs and I think capacity is reallocated over time; for example, a 22nm fab might be converted to 10nm to adapt to changing demand. They can eventually get down to just one 22nm fab which won't require as much demand to fill.

To be fair, there are limits on how far and how fast one can upgrade a fab and upgrades are planned out many years in advance. Equipment for the smaller nodes tends to be larger and more complex and involving more steps; this requires more space. Additionally, ASML only makes around a dozen of the EUV machines each year and they are booked years in advance.

What percentage of the equipment changes between nodes ? Is it mostly the lithography and inspection tools ?

Intel fanning for themselves is more than twice the business of TMSC fanning for many others.

https://www.extremetech.com/computing/317769-intel-tops-semi...

What hurt Intel was missing the mobile CPU boat. If Intel wasn't so braindead in management, they probably could have won a substantial segment of mobile.

Because of mobile, TSMC had so much revenue to advance their process and catch/surpass Intel. Heck, Apple's M1 is probably due to the same reason in the Apple ecosystem, they couldn't win the iPhone either, and IIRC various "what's wrong with Intel" threads stated many of the best Intel engineers went to Apple.

> Because of mobile, TSMC had so much revenue to advance their process and catch/surpass Intel.

Intel never been chasing pure revenue. They were always conscious of their profitability, to please sock market analysts.

Agreed, that's what happens when the MBAs infect your company.

Same with Medtronic, a medical engineering company gets overridden with MBAs and now the only innovation they do is acquiring other companies.

each iterative cycle requires exponentially more r&d, making it prohibitive for that one firm to catch up

I’ve been told it helps protect Taiwan from prc aggression through its allies dependency on the fabrication continuing.

Like the myth that China's GFW is to protect local business, this is just retrospective fairytale.

TSMC succeeded because of the chip and high-quality talents on the island. It was not setup to become a high-value precious.

And the logic also easily break down, if anyone bother to think about the consequence. I.e., put a high-value target in a contended geographical location, just makes the conflict more desirable for the aggressor.

Sure the protector would be more concerned if the target is lost, but the value of acquiring that target suddenly becomes way more attractive, in comparison to the potential loss to the protectors.

The value of the target is probably irrelevant. What matters is probability of success. Any Chinese will be decided by expected loses and gains. This is probably why IS continues to arm Taiwan as it is the biggest deterrent.

I am replying to the idea that TSMC is setup to be valuable to US, such that Taiwan get more protection.

You are mistaken. This was a deliberate and overt tactic by the Taiwanese government. It's been heavily discussed in geopolitical circles for decades.

TSMC is not something that can be "acquired" by military force. That's not how things work. If China does follow through on their threats to take the island by force, they won't be able to simply operate TSMC as if they were new management. The organization itself will cease to exist, along with all the institutional knowledge necessary to duplicate their success, even assuming China was able to secure the fab lines without damage. Likewise all customer relationships become essentially void, and the finely tuned money machine that sustains TSMC's intense R&D ceases to exist as well.

Looks like the Taiwanese government forgot their tactic - because China is hiring away TSMC engineers to bolster China's chip manufacturing R&D.

https://asia.nikkei.com/Business/China-tech/China-hires-over...

The situation is much more complex than "forgot." Since the 80s KMT and several allied parties have favored eventual unification with the mainland. Try reading up on the Pan Blue Coalition, or just skimming the wiki page on Taiwanese political parties.

It gets tiresome how flippant hn is to throw out these narrative based explanations that are pure invention of armchair imagination, rather than like spending literally 5 minutes to understand the very basics of what people are attempting to comment on.

What would the alternative be?

Slowing all progress down so everyone can keep up?

I assume keeping the parts cool becomes a big issue. You get twice the amount of heat to dissipate within the same area. Or is there some workaround?

No Workaround AFAIK, the idea right now is not to stack compute die on top of each other but other things like SRAM or parts that uses little energy for both cost and space efficiency.

The goals here are kind of sideways from heat dissipation.

* Different parts from different processes — for example DRAM and CPU need different processes, so it doesn't work to put them on the same chip — this integration is a step closer than on-package.

* Yield — if you really want sixteen cores, and you're pushing the edge of your process, your yield for perfect eight-core chiplets will be dramatically better than for a sixteen-core chip or chiplet.

But yeah concentrating all that into multi-chip chip will also concentrate the heat.

> Or is there some workaround?

Synthetic diamonds, heatspreaders in the package

*Isotopically pure diamonds

The work around would be microfluidic coolers, which aren't production ready yet, I believe.

It is like national security matter and relatively inexpensive when talking about nation's budget scale. Why countries don't make their own chips? Should this be in the realm of public companies just like healthcare or energy in some countries?

The US does recognize the strategic importance of semiconductor manufacturing and spends billions on it each year. Many devices for markets like aerospace and defense are only manufactured domestically. That said, the US has a defensive pact with Taiwan and to attack the island would be to start World War 3.

It's not really inexpensive in terms of a budget when you look at the nations that could benefit from the investment as a national security concern. Look outside the US nuclear umbrella and NATO to nations not friendly with the PRC - who among them is wealthy enough to start a chip industry that can compete?

There is no defensive pact with Taiwan. There will be no WW3 should the PRC decide to go for it.

Whoa! Blew through my marketing gibberish budget less than one paragraph in.

Could this be used for the next generation of HBM?

Not much. HBM will not gain much from going inside the chip pachage over existing interposer solution.

Though, you may decide foregoing the interposer. This way, money go TSMC instead of interposer maker, which was already likelly been TSMC.

I wish all this technology were more accesible. but they gotta protect their trade secrets (I guess...).

I've heard about something called 'dycryl process' (in chip photolitography) but it is not the kind of info that it's easily found, nor easily understood even if I could google for it.

>I wish all this technology were more accesible. but they gotta protect their trade secrets (I guess...).

open source their manufacturing process? everyone can get EUV machine and all the components to make chip. you figure out the better yield part or packaging like TSMC

They could give you all the recipes and steps, but a newcomer isn't going to make it due to the needed experience and expertise.

Reply on news.ycombinator.com