News TSMC explores using 510x515 mm rectangular silicon wafers — tripling the usable area of current 300mm diameter tech

[Admin]

Nikkei says TSMC is researching rectangular substrates for advanced chip packaging.

TSMC explores using 510x515 mm rectangular silicon wafers — tripling the usable area of current 300mm diameter tech : Read more

 

[edzieba]

Anyone remember 450mm wafers? Can't see why this would fare any better.

 

[gondor]

How do they grow rectangular crystal?

 

[usertests]

Anyone remember 450mm wafers? Can't see why this would fare any better.

It will fare better if it fares better.

How do they grow rectangular crystal?

Carefully.

 

[KnightShadey]

How do they grow rectangular crystal?

Could be horizontal melt, vs the typical cylinder. It's been years since I thought of this stuff, but that would make sense for making a big thick flat slab/sheet vs vs slicing a long cylinder.
Even with crystal alignment considerations handled slicing a cylinder lengthwise would still only yield X x ~300mm at it's widest point and only allow a few sheets at that width progressively narrowing.

(edit 4th time's a charm, direct link to imgur not using the insert tab)

https://imgur.com/a/osEtb5j

View: https://imgur.com/a/osEtb5j


https://www.researchgate.net/profil...l-Bridgman-HB-crystal-growth-process-in-a.png

So to me it would have to be some horizontal process vs the tradition vertical to achieve that shape.

I wouldn't be surprised if much of this desire for bigger wafers is likely precipitated from the recent push of designs past the reticle limit resulted in making larger and larger chips and not being able to fit enough within 300mm diameter with large amounts of waste, whereas a rectangle would offer more flexibility, and likely even bring down failure ratio per wafer, just by having more wafer to put more chips on (assuming quality control gets you a similar enough starting point)

Just some guesses on my part, could be melted pop-rocks for all I know.

(edit for some reason the image is blocked, even though it shows up in preview (usually if unacceptable it gets rejected in preview? 🤷🏻‍♂️)
Fixed by direct link to imgur, ugh!

 

[DavidLejdar]

Anyone remember 450mm wafers? Can't see why this would fare any better.

This blog post by Paul McLellan, explains it quite well. The video at the bottom there, is also interesting.

Summarized, the transition i.e. to 450mm requires a lot of investment. Like entire fabs need to be replaced. The equipment makers also need to be able to handle it. And the return of investment for previous transition wasn't as fast as hoped for.

There is also the version of How TSMC killed 450mm wafers for fear of Intel, Samsung. It could mention though, that in 2013 TSMC had less than half the yearly revenue as Intel did, and unlike Intel and Samsung, TSMC doesn't have its own retail products, and their fab capacity may have been plenty for the orders by other companies, while it did make sense to focus more on refinement, such as moving from 22nm process in 2012 to a lot less these days. (Or in other words, that a GPU these days can do a lot more than a GPU from ten years ago, at almost the same size, that has a lot to do with refinement.)

And it is rather only now, that there is a sort of necessity to move on from 300mm. As mentioned in the Nikkei article, only 16 sets of the B200 can be had from a single wafer, compared to 29 sets of H200 and H100. That likely means, as pointed out, way more waste (due to the "larger rectangles"), and possibly even running out of space, as products for datacentres do not necessarily have to fit under or on a table. In that context, it seems reasonable to look into doing it differently, than to just push for 450mm, at a cost per wafer likely quite exceeding the nearly $17,000 for a 300mm wafer at 5nm, mentioned several years ago.

 

[DaRAGingLunatic]

Anyone remember 450mm wafers? Can't see why this would fare any better.

The industry was going to try moving to 450mm wafers, well at least intel was… but TSMC decided not to as all the tooling was for 300mm wafers.

Without wafer size increasing, costs have gone up per die due to better nodes costing more per wafer due to fabs costing more.

If we can get larger dies then perhaps the costs per die can be reduced making it more economical gor consumers.

 

[Kondamin]

This blog post by Paul McLellan, explains it quite well. The video at the bottom there, is also interesting.

Summarized, the transition i.e. to 450mm requires a lot of investment. Like entire fabs need to be replaced. The equipment makers also need to be able to handle it. And the return of investment for previous transition wasn't as fast as hoped for.

There is also the version of How TSMC killed 450mm wafers for fear of Intel, Samsung. It could mention though, that in 2013 TSMC had less than half the yearly revenue as Intel did, and unlike Intel and Samsung, TSMC doesn't have its own retail products, and their fab capacity may have been plenty for the orders by other companies, while it did make sense to focus more on refinement, such as moving from 22nm process in 2012 to a lot less these days. (Or in other words, that a GPU these days can do a lot more than a GPU from ten years ago, at almost the same size, that has a lot to do with refinement.)

And it is rather only now, that there is a sort of necessity to move on from 300mm. As mentioned in the Nikkei article, only 16 sets of the B200 can be had from a single wafer, compared to 29 sets of H200 and H100. That likely means, as pointed out, way more waste (due to the "larger rectangles"), and possibly even running out of space, as products for datacentres do not necessarily have to fit under or on a table. In that context, it seems reasonable to look into doing it differently, than to just push for 450mm, at a cost per wafer likely quite exceeding the nearly $17,000 for a 300mm wafer at 5nm, mentioned several years ago.

In 2008 venture capital dried up shifting everything on tsmc is a bit silly.
It probably didn’t help they quit half way trough but at the time intel could have just put its entire weight behind it and done it them selves were it not for the global financial crisis they probably would have just done that.

Just one of the many projects intel stepped out of because of penny pinchers, which had they persevered would have put them in a far better possition

 

[jp7189]

How do they grow rectangular crystal?

Perhaps just grow a massive cylinder and cut it down to a rectangle? It's probably cheaper overall to waste the edges at that stage rather than after the fab stages.

 

[DS426]

That's nVidia's and anyone else's problem if they are growing their chips that large. Produce smaller -- yes, lower performance -- chips but have less waste and more GPU's, etc., ultimately resulting in better costs and more usable die being put into the world. Instead, these chip makers want to have the biggest, loudest flagship product in the room as if it's a showing off their male appendage size.

This also relates back to the soaring TDP problem (or at least what's a problem IMO).

 

[KnightShadey]

Perhaps just grow a massive cylinder and cut it down to a rectangle? It's probably cheaper overall to waste the edges at that stage rather than after the fab stages.

You'd want to keep the edges until dicing to reduce fragility.

However, to accommodate for a 500mm square (rounding down and using Pythagoras) you would need a 700+ mm diameter cylinder to be able to cut that width, and they struggled going to 450mm. 700mm would be more than doubling the current diameter.

 

[KnightShadey]

That's nVidia's and anyone else's problem if they are growing their chips that large.

True, but if nVidia (or anyone else like intel, Samsung, GloFo, etc) solves the problem, then why do they need TSMC?

TSMC want's to solve the problem of their Trillion dollar client, for the very big manly product that makes them all their money and valuation so they don't use $100-200 billion of that value to create their own foundries.
Also, if TSMC solves the problem, then they can sell the solution to AMD, Tim Apple, Google, intel, M$, etc. as well as nVidia.

Big wafer also means many more little chips on a single wafer as well, might even be better economies of scale if a horizontal process reduces waste while producing more usable surface area.

If it was AMD valued at $3T , it would be ironic for them to get back into the foundry business (even buying struggling GloFo back for a mere 10-15B) coming full circle 15 years later. 🤔😉