[DylanBohlender]

Here are a couple good articles explaining it.

https://www.extremetech.com/extreme/227720-how-intel-lost-10...

https://semiwiki.com/semiconductor-manufacturers/intel/28919...

The TL;DR is that Intel has always been a vertically integrated shop (meaning that they usually fab and design their own chips), and that is starting to bite them because pure-play foundries are improving their tech at a faster rate.

Intel has been unable to keep up with process advancements in their foundries, and that has led to pure-play foundries like TSMC taking massive market share. As chips get smaller and smaller, Intel has failed to keep up. They can only do 10nm for their mobile stuff and 14nm for their desktop stuff, whereas fabs like TSMC have been in 7nm territory for a while now and are moving into 5nm territory.

[ckemere]

Just skimming the first article, the TL;DR that I took away is that TSMC can get revenue from their old foundry nodes for much longer than Intel can. So the issue is not how fast TSMC improves, but more that Intel has to invest proportionally much more to keep up.

[dragontamer]

Note: the I/O chip on AMD Zen 2 / Zen 3 is a 14nm GloFo chip. Only the CPU-cores ("Zepplin" maybe, or whatever they call them now) are 7nm TSMC.

So AMD's strategy also leads to lower fabrication costs: because they can make a far cheaper 14nm chip to handle the slower portions of I/O (talking to RAM, or PCIe), while the expensive 7nm parts of TSMC are used only for the cores / L1 cache / L2 cache / L3 cache.

[zinekeller]

Small nitpick: if I remeber correctly, AMD now uses 12nm process from GloFo now instead of 14nm (due to PCIe 4 requirements).

[dragontamer]

If so, it hasn't changed since Zen 2. IIRC, AMD is on record saying that Zen2 / Zen3 have the same I/O chip, and that only the 7nm stuff is changing.

Zen 2 has PCIe 4, so you could very well be correct.

[Teever]

Is there any indication that Intel will move to this chiplet design? It seems like it's a no brainer and it's been proven to work by AMD.

What's keeping Intel from adopting this same practice? is it IP?

[dragontamer]

Intel has a competitor to chiplets, called EMIB, based off its Altera purchase. EMIB is pretty cool, but has only been deployed in a small number of situations so far (There was a Xeon + FPGA chip Intel made. There was the hybrid Intel+AMD chip, and finally the new big.LITTLE clone chip that Intel merged an Icelake + Atom core together). I don't know why Intel hasn't invested more heavily into EMIB, Foveros, and other advanced-packaging technologies... but Intel is clearly working on it.

Intel can do it, they just haven't decided to do so yet. They have the tech for sure.

Its simply a matter of priorities. Its not so much that Intel "isn't" investing into it, its arguable that Intel just hasn't invested "enough" into it.

AMD went all in: they literally bet their entire company on advanced-packaging, with AMD GPUs using an active interposer with HBM2, and now Zen-chips taking a chiplet strategy. And to be fair: AMD had to do something drastic to turn the tide.

We're just at a point where AMD is finally reaping the benefits of a decision they made years ago.

--------

If I were to take a guess: Intel was too confident that they could solve 10nm / 7nm (or more specifically: EUV and/or quad patterning), which would have negated the need for advanced packaging.

AMD on the other hand, is fabless. They based their designs off of what TSMC was already delivering to Apple. Since TSMC leapfrogged Intel in technology, AMD can now benefit from TSMC (indirectly benefiting from from Apple's investments).

Intel's failure bubbled up from the fab level: Without 10nm chips, Intel was unable to keep up with TSMC's performance, and now AMD is advancing.

----

AMD's strategy just works really well for AMD. AMD is forced to keep buying chips from GloFo (which are limited to 14nm or 12nm designs). All of AMD's decisions just lined up marvelously: they fixed a lot of issues with their company by just properly making the right decisions in a lot of little, detailed ways. A happy marriage of tech + business decisions. I dunno if they can keep doing that, it almost feels lucky to me. But they're benefiting for sure.

AMD took something that seemed like a downside (the forced purchase of 12nm or 14nm chips, even when 7nm was available), and turned it into a benefit.

[AnthonyMouse]

> They based their designs off of what TSMC was already delivering to Apple. Since TSMC leapfrogged Intel in technology, AMD can now benefit from TSMC (indirectly benefiting from from Apple's investments).

Did Apple actually buy a significant stake in TSMC or are you just referring to the fact that Apple is one of their large customers along with Qualcomm, Nvidia, Huawei (until recently) etc.?

[dragontamer]

> Did Apple actually buy a significant stake in TSMC or are you just referring to the fact that Apple is one of their large customers along with Qualcomm, Nvidia, Huawei (until recently) etc.?

I'm talking more like the later: all of these companies (Apple, Qualcomm, NVidia, etc. etc. and of course, AMD) are effectively pooling their money together to fund TSMC.

I don't mean to single Apple out as if they're the "only" ones funding TSMC's research. (And I can see how my wording earlier mistakenly can be interpreted in that manner. I was careless with my wording). Its more of a team effort (although Apple does seem to spend significant amounts of money trying to get first-dibs on the process).

[nolok]

For better or worse AMD has always been much more willing to quickly throw everything out to go with a completely new design paradigm. Something it was a massive bust like X2 or Bulldozer. Sometimes it's the Athlon or the Zen.

The only time Intel really did that was with the end of the P4 and frankly even then they waited as long as they could before doing it. The rest of the time it's all carefully planned stepped increase and safe design change.

Both have their advantages, but for your question it means Intel will have to take a leap they clearly don't like taking.

[rasz]

Historically most successful AMD designs came from the outside. 286 was Intel clone, Am386 Intel microcode copy. Am486 lagged by one year, offered lower performance and was still developed using Intel IP. K5 first 100% AMD design, slow and late, competed in the bottom low end, considered a failure. K6 100% external design by NexGen, let AMD move up to middle of the market. K7 designed by DEC Alpha team and manufactured in Motorola partnership, great success.

[avmich]

> The TL;DR is that Intel has always been a vertically integrated shop (meaning that they usually fab and design their own chips), and that is starting to bite them because pure-play foundries are improving their tech at a faster rate.

Are you saying that vertically integrated companies are inherently disadvantaged because pure-play companies have a bigger list of orders and can spend more on evolving technology?

There is advantage to be vertically integrated, namely a better global optimization across domains, like between design and manufacturing. Would be interesting to know why it's not enough - if it's always not enough. If not always, then this particular case needs more reasons.

[ethbr0]

The balance likely comes down to volume. Specifically, total global chip sales.

If Intel builds Intel (I believe their contract fabing is a rounding error?), then they're directly tied to Intel chip sales.

This sets up a potential death spiral. Intel misses a process node deadline, Intel's products are uncompetitive, Intel sales decrease, less demand for Intel fab, less money for Intel fab improvement.

Intel can temporarily paper over this by shifting money from other areas of the company, but it's not a good path to be on.

Conversely, as you might expect, if Intel sales are increasing then the opposite, virtuous cycle holds.

So essentially, Intel's fortune is tied to the Intel_sales : (global_sales / number_of_leading_edge_non-Intel_fabs) ratio.

And with regards to that, two huge things happened in the marketplace recently: (1) mobile chip sales explosion, (2) GlobalFoundries exiting leading process race.

If Intel hadn't been screwed by a process engineering miss, longer term trends would still have hit them hard.

[CountSessine]

This is the case with every foundry, though. It’s the reason GloFo isn’t competitive anymore, for example. What’s more interesting is the physics reason that they tripped up the last generation - what did TSMC do right that Intel did wrong? What bets were made? Which ones paid off?

[Pet_Ant]

I believe the issue was that Intel was leaning harder on EUV trying to make it and burying the competition instead of a more cautious approach by TSMC. Zen 3 is finally using some EUV layers whereas I believe Intel already wanted to use EUV heavily in their "10"nm process.

[wmf]

Nope, Intel 10nm does not use EUV. Zen 3 is made on TSMC N7 (not N7+ as rumored) which also does not use EUV.

[ethbr0]

I feel like the business models are slightly different when you're a contract foundry vs integrated though.

The latter depends on your market share. The former only depends on the total market.

[jandrese]

I think the idea is that the open fab shops just have a hell of a lot more work than Intel and greater economies of scale. Making zillions of mobile chips means big money even if the profit margin per chip is smaller. This in turn means more R&D and eventually they overtake the company that only fabs their own desktop processors and chipsets.

In the long run Intel screwed themselves over by not leasing out fab time to other companies. They put themselves in a niche in an industry that is naturally dominated by the largest player. And it's extra embarrassing that they did so because they knew very well how important it was to be the biggest--they were for a few decades!

Maybe Intel could have held on longer if they had a successful mobile chip to stuff into billions of smartphones, but their mobile efforts were short lived and seemed to be treated with disdain by the management. The first product kind of sucked and instead of sticking with it and improving they just threw in the towel, both on mobile processor and the baseband chip. An embarrassing misstep for a company as big as Intel.

[trhway]

>In the long run Intel screwed themselves over by not leasing out fab time to other companies.

sounds like Google who also treated their cloud/fabric computing as a competitive advantage to be kept to themselves and as result they missed the cloud business.

>their mobile efforts were short lived and seemed to be treated with disdain by the management.

classic. Low margin high volume future usually can't survive in the shadow of the high margin cash cow of yesterday that is still being milked.

[simonh]

I think the point is if TSMC develops a new node they will find customers for it, and they compete with other foundries on node.

With intel their foundries only have internal customers, who can only go to their internal foundries. There’s no competitive pressure on the foundries, and if they are ready early that’s wasted capacity. So so capacity and technology planning is based on a common road map to meet the needs of their slowest (er, I mean only) customer.