Hacker News Reader

Comment by datadrivenangel | Hacker News Reader

These chips are still sent to Taiwan for packing, so it's a good step but not a complete step.

Until 2027, yes.

https://www.tomshardware.com/pc-components/gpus/tsmc-is-repo...

"TSMC does not have an advanced packaging facility in the U.S., and its partner Amkor will only start packaging chips in Arizona in 2027. As a result, Blackwell AI silicon produced in Arizona will need to be shipped back to Taiwan for final assembly, as all of TSMC's CoWoS packaging capacity remains in Taiwan."

Given that there may be a 25% chance that China invades Taiwan by 2030, having the ability to package SOTA chips in the US by 2027 seems "soon enough".

where did you get that number from

There's a window where China will have it max capability to invade for the next few years. After that their population is going to start shrinking and every year will be harder than the next to invade.

The population of 18-30 year old males is generally what matters for an invasion and China has been shrinking that for a long time. The rest of the population can plan the invasion, but they rarely actually do it. (a few countries also invite young females to an invasion, but that is not normal)

But they don’t have energy independence or food security yet, which is kind of a hard requirement for an invasion.

There’s not enough rail lines and gas pipelines from Russia to feed them with significant quantities of fossil fuels.

Imagine how bad Russias invasion of Ukraine would’ve been without energy independence and food security. The invasion of Taiwan is an order of magnitude more difficult, and Taiwan now has the recipe for how to knock out the entire naval fleet of a more powerful nation (see how Ukraine has essentially incapacitated Russia in the Black Sea).

I would expect an invasion to prompt the US navy to put up a blockade, disrupting China's oil supplies and generally making it very hard to keep their economy going. Admittedly, Trump is a wild card; he's random enough that it is hard to be sure what would happen.

I do not think China could survive a blockade.

Huh? China will get all the energy it needs from Russia. Plus China is the leader in renewables production.

Not a chance. Russia has no way to deliver oil in large quantities to China. New pipelines, if they could even build them, would take many, many years.

If renewables was adequate they would not be importing vast amounts of oil.

Except TW TFR even worse than PRC TFR, and ultimately scale effect takes over - PRC with crippled TFR still generates about as much male new borns per year than TW has men 18-40 total. PRC still on trend to generate 3-4x more MEN than US projected to add population per year, incidentally around the same as active duty military... having enough bodies is not going to be an issue for decades. Having enough nukes is.

I'm not arguing against that at all. Just that if the PRC wants it's best chance, the clock is ticking. It becomes more costly the longer they wait.

I disagree, bodies is not limitting factor for PRC, it also becomes cheaper to wait for TW specifically because TW male 18-40 is set to decline = less kill bots / occupation force needed. Attacker:defender ratio (i.e. commonly 3:1) = every defender TW loses due to demographics, PRC with same TFR will come out significantly ahead, will need less enforcement:civilian ratio for occupation.

But ultimately, it's about hardware+industry - current trend = regional force balance shifting in PRC favour vs US+co every year with no end in sight. PRC better off accumulating capabilities at scale, not just regional, but global (i.e. prompt global strike) and increase autarky (less net population + more electrifcation = more calorie + energy security). All trend incentivizes waiting and building.

TLDR waiting and building becomes less costly (or rather less risky) to pursue PRC's ultimate strategic goals associated with TW scenario... displacing US posture out of east Asia and perhaps hitting CONUS infra at scale as response to US intervention. The latter part is key, there are important stretch goals to TW scenario that secures PRC geopolitical interests for 50-100+ years. It's much more important to be able to tackle those "costly" scenarios "cheaper", where cheaper is also relative to making intervention much more expensive for adversaries, i.e. PRC "winning" hand in TW scenario is to show US posture in east asia not sustainable, and CONUS (including TSMC Arizona) not defendable.

You know where

I'm not sure where GP's 25% comes from. But there have been various assessments that China intends to "reunify" with Taiwan by 2030. [1] Xi Xinping has also instructed the PLA to be prepared to invade by 2027. [2]

If you then ask yourself whether China would rather invade during the Trump administration (with its tendencies towards isolationism and "deal making") or roll the dice on a subsequent U.S. administration, you might find yourself thinking that the odds actually seem considerably higher than 25% that this could happen in the next four years.

To the extent that this narrative comes via the U.S. intelligence/defense community, one has to assume that it may biased towards exaggerating the threat. I for one hope that is the case, since I do not want to see a U.S.-China conflict any time soon. At the same time, I unfortunately don't think it's likely to be completely baseless.

[1] https://media.defense.gov/2023/Apr/24/2003205865/-1/-1/1/07-...

[2] See, e.g., https://cimsec.org/the-maritime-convoys-of-2027-supporting-t...; https://thehill.com/policy/defense/4547637-china-potential-t...

Metaculus - this is the median prediction.

Would be interesting if China uses drones with technology from Taiwan to invade Taiwan.

what is involved in the packaging process ? I believe they don't ship fully assembled chips to Taiwan only to be put in a pretty box ?

I believe packaging in this context means taking the raw silicon dies and assembling them into a package which can be soldered onto a PCB (or put in a socket, but Apple doesn't socket anything).

I'm making an educated guess but probably the cutting of chips from the wafers, placing them into the appropriate ceramic socket types (DIP, BFGA, SMD etc), soldering the line wires from chip to pin, encasing the chip, etc.

> DIP

I am happily imagining opening a recent Apple device and seeing 74 gates with through holes in green PCBs, with an Apple logo made in soldering lead marking in the corner of the board.

I think "packaging" here refers to the process of putting the silicon die in its plastic casing and connecting the die's pad to the case's pins, see https://en.wikipedia.org/wiki/Integrated_circuit_packaging

"Packaging" in this context means taking the wafer of compute die (made in Arizona), dicing it up into individual die, mounting it onto a silicon interposer (an even bigger die, no idea where that's made, but probably taiwan) along with a bunch of HBM die, then mounting that Si interposer on a somewhat larger, very fine-pitched circuit board ('substrate') that is essentially a breakout for power and high-speed I/O from the compute die. That thing is the packaged 'CoWoS' system, where CoWoS==Chip-on-wafer-on-substrate, that eventually gets attached to a 'normal' PCB.

This sounds like a complex procedure. Are there currently alternative packaging facilities that could do this work, if Taiwan were locked into kinetic war?

What I've always wondered was, how is it possible to do this process (or well, the less advanced version of it, for smaller/older chips) cheaply/at massive scale, for those ICs that cost a few cents in bulk?

Like, scaling wafer (die?) production to insanely low costs makes intuitive sense. The input is sand, the process itself is just easily-parallellizable chemistry and optics, and the output is a tiny little piece of material.

But packaging sounds as though it requires intricate mechanical work to be done to every single output chip, and I just can't wrap my head around how you scale that to the point where they cost a few cents...

https://www.youtube.com/watch?v=-egYoxajTz0

Believe it or not, sending them overseas just to be put in a box actually can be cost-effective. Like with those pears: "grown in Argentina, packaged in Thailand, sold in UK" https://www.birminghamfoodcouncil.org/2022/01/16/part-i-pear...

How does this make any financial sense?

This is how most modern supply chains look like.

Plus, chips are small in size and cost a lot so you can fit a lot in a container. Per unit shipping costs probably come out to be pretty low. Especially when compared to the political costs and risks associated with not onshoring.

> you can fit a lot in a container

Guys these are microchips on wafers. You can put a million dollars worth in your jacket pocket. They aren't being shipped in containers.

Marine shipping is just about the most fuel efficient way of moving things between any two places, by a lot. A 100,000 dwt ship can get 1050 miles per gallon per ton of cargo. It takes about a teaspoon full of fuel to move an iPhone sized device across the pacific when I ran the numbers last.

Interesting. Could you give a brief description of how you got that number? Eg. what factors were considered.

Those numbers match what comes up with a quick search:

https://www.extension.iastate.edu/grain/topics/EstimatesofTo...

That study uses 1,043.4 mpg for the fuel economy of a 100,000 dwt ship.

Videos of transportation ship engines are cool. Each cylinder is wide enough for a person to lay down inside it.

https://youtu.be/G0eMyA388bE

To ship things to/from these fabs by sea you have to add the cost of shipping by truck between Phoenix and (presumably) LA. Not sure how big of a difference that makes.

Chips are small, so one truck once a few days may suffice.

Airplanes. They use airplanes. We are talking about microchips here, possibly the highest dollar per gram substance that exists on the planet.

The interest you'd pay just losing a couple days in transit time would exceed the cost of purchasing a dedicated private jet and the crew to fly it.

Imagine the insurance...

A semi truck carries +- 15 tons of cargo and gets an average of about 6 MPG, so about 90 MPG/ton.

Trains are pretty efficient as well.

I'm suprised they can't ship (flat) packaging that could be used in Arizona with a simple assembly line.

If they had that packaging design then for this to make financial sense the two way shipping (and loading, unloading, custom clearance etc) would have to be less than shipping the packaging, the setup cost per unit cost of putting the chip in a box

Wait, wait. In the context of semiconductor manufacturing packaging does not mean what you think it means. It is not putting the product in a paper box.

It is about cutting the wafer into individual chips, wire bonding the silicone to pins, and covering the whole thing with epoxy.

Here is a video which explains it better: https://www.youtube.com/watch?v=7gg2eVVayA4

It would be indeed crazy if they would ship the ready chips to Taiwan just to be put in a paper box.

basically the input of the process is a wafer which looks like this: https://waferpro.com/wp-content/uploads/2016/08/Patterned-Lo...

And the output of the process is something which looks like this: https://res.cloudinary.com/rsc/image/upload/b_rgb:FFFFFF,c_p...

The packaging in this context is not wire bonding but CoWoS - chip-wafer and wafer-wafer bonding.

You are correct. I was just illustrating what kind of processes belong to the umbrella term "packaging" in the context of semiconductor manufacturing. Was not talking about what particular process are missing from the Arizona facility.

But you are right on that it is CoWoS which is the missing ingredient.

Thank you for clarifying!

You seem to be confusing the term packaging...it is not the box, it is how the chips are assembled together to make the final product.

https://www.youtube.com/watch?v=-egYoxajTz0

Dont downvote the guy for not knowing this very specific definition of 'packaging'

Right, but he assumed he knew a technical term when he didn't, which is unwise.

The machines and processes needed to package the individual integrated circuits are fantastically expensive but the margins are so low in that step that it's only profitable at massive scales.

So you put the fantastically expensive machines near where most of the customers are and most of the customers are in Asia.

Works the same way with fiber optic cables. Making the long skinny bits is hard and high-margin. Actually turning them into cables is easy and low-margin.

So Corning makes huge spools of fiber optic cable in Arizona, North Carolina, and New York (I think) and ships it off to Taiwan and China where it is made into the cables that you plug into stuff.

These are literally microchips. Tens of thousands of dollars of value in each gram.

Shipping cost is fundamentally irrelevant, you can put $100MM worth on a direct flight and have room left over for your family and friends.

Your overall point is probably right, but "tens of thousands of dollars of value in each gram" seems like an exaggeration. How much does one CPU weigh?

Order of magnitude it's within range. A single wafer for something higher end is worth tens of thousands of dollars. So whatever that weighs. It's not much.

But surely a wafer weighs more than a gram, no?

Yes. One source I just double checked says that a 300mm silicon wafer weighs 840 grams, which seems to be in the right range anyway.

I see that source too (universitywafer.com) which is sucked into the AI slop panel on Google.

A 300mm wafer at 0.775mm thickness that weighs that much would be a density of 15g/cm³: midway between lead (11) and gold (19). Silicon is only 2.3g/cm³.

A 300mm wafer should be much closer to 100g than 1kg. Another page (on the same website!) says 125g, which is pretty dead-on 2.3g/cm³.

Depends on the chip. A typical consumer processor is 100-200 mm². Wafers are about 1mm thick (actually a bit less, but close enough) and silicon has a density of 2300kg/m³. So it's 2.3mg/mm² of wafer area. The chip is then 0.23-0.46g. if it's high end and worth $1000, then the wafers are $2-4000 per gram. Probably the low end of that since the fancier chips are usually the physically bigger ones.

However, low-end processors will be worth radically less ($100 CPUs aren't 10x smaller) and things like top-end FPGAs will be worth substantially more. An Agilex 7 can be $40,000 and if it's got less then 4 grams (around 1700mm², or 41mm*41mm), the wafer is worth over $10k/gram. The entire chip is 56x56, and from the package drawing, the die appears to be around 30x30, so it exceeds the threshold.

This is assuming all the value of the retail price is the main wafer, which is also not true.

Thousands per gram, certainly seems possible, though I doubt it's even that on average across all chips, and considering that there is non-wafer content in the price. Tens of thousands is probably pushing it. However, it's certainly likely to be rather more than gold in terms of specific value ($100/g, ish). Harder to fence, though.

https://www.freightwaves.com/news/dsv-unveils-details-for-ne...

^certainly there's activity in that space

What does "packaging" mean in this context? I'm a total n00b when it comes to chips.

What comes to my mind is wrapping a piece of electronics in some bubble wrap and cardboard, which doesn't sound that hard...

Reply on news.ycombinator.com