Why wasn't TSMC doomed then? And why couldn't Intel turn around now like TSMC did?
Intel looks nothing like other vertically integrated CPU businesses that divested from manufacturing in the past, it's still making a lot of money and gross margins are still healthy (and still higher than AMD and TSMC at the moment!) although clearly taken a hit from the mid/low 60s where they typically sat down to mid/low 50s in the past 3 years. Manufacturing silicon is still a profitable business for them.
Intel also has by some important metrics, the best performing CPUs available despite competitors being on TSMC and Samsung. I don't argue the others aren't better in other equally valid metrics. It was actually much more difficult to make that same statement about competitors back when Intel had a clear silicon tech advantage.
So definitely Intel is a leading edge node manufacturer if others were in the past.
In that sense it was almost irrelevant to the rest of the market what node Intel had achieved, since nobody else had access to it. Similarly, because it was core to their business model, the foundry business itself was not very cost sensitive; as long as the process tech was exclusive and enabled them to make the world's best processors, they could extract princely sums for the final products.
This has fundamentally changed in a way that will not revert. First, we have reached the "end of Moore's Law": process improvements will still continue, but no longer at a pace such that process leadership guarantees performance leadership. Second, without this guarantee of performance leadership, Intel's entire manufacturing operations are having a crisis of purpose. Their manufacturing operations are much less cost-effective than TSMC because that efficiency was never really a major driving force in the past (and to some extent was at odds with the goal of innovation and leadership).
Without the serious prospect of regaining a performance advantage, Intel will probably not be investing the kind of money needed to displace TSMC in process leadership. Instead the focus is going to be on making sure their existing fabs are not stranded assets and figuring out how to effectively compete on price/performance without a persistent process advantage.
That doesn't follow. Intel has been massively relevant to the semiconductor manufacturing market for 50 years. Arguably the biggest single influence and the one which gave rise to (high performance) foundries in the first place by driving vertically integrated competitors out of business.
> This has fundamentally changed in a way that will not revert. First, we have reached the "end of Moore's Law": process improvements will still continue, but no longer at a pace such that process leadership guarantees performance leadership. Second, without this guarantee of performance leadership, Intel's entire manufacturing operations are having a crisis of purpose. Their manufacturing operations are much less cost-effective than TSMC because that efficiency was never really a major driving force in the past (and to some extent was at odds with the goal of innovation and leadership).
I don't follow this either. What fundamentally changed? Foundries have been around for almost as long as the semiconductor industry (UMC founded in 1980), and have been huge players in the market since around 2000.
> Without the serious prospect of regaining a performance advantage, Intel will probably not be investing the kind of money needed to displace TSMC in process leadership. Instead the focus is going to be on making sure their existing fabs are not stranded assets and figuring out how to effectively compete on price/performance without a persistent process advantage.
Intel has publicly stated their plans to build new fabs, invest in new nodes and close the gap with competitors in the next few years so that doesn't appear to be their focus or assessment of their prospects. Whether they actually achieve that or not remains to be seen, but if you are going to say they can't, then you're left explaining how TSMC could -- back as recently as 2015 or so you can see articles wondering whether TSMC will be able to close the gap with Intel. Nothing suddenly changed between 2015 and now, Moore's law wasn't okay then and suddenly just stopped in 2018 when Intel started falling behind, so I don't think you can just handwave about that to explain it away.
Well, sure: Intel pioneered FinFET, cobalt, even 300mm. TSMC and the rest of the foundry world basically followed their lead. But as any foundry customer, it didn't matter what Intel's state of the art was because you couldn't pay any amount of money to obtain anything other than a final Intel product.
> I don't follow this either. What fundamentally changed?
Limiting physics. It is very hard to continue to make devices smaller at historical rates. Nobody is keeping up with the historical 2-year node cadence any longer. And the "nodes" now are marketing smoke and mirrors, anyway; they don't actually mean that device density has doubled.
> Intel has publicly stated their plans to build new fabs, invest in new nodes and close the gap with competitors in the next few years so that doesn't appear to be their focus or assessment of their prospects.
They cannot publicly pull the plug for two major reasons:
1) They need to attract customers for their move into the foundry space (again, their fabs becoming stranded assets are a nightmare scenario)
2) They're hopeful of getting US government subsidies in order to continue competing on a basis other than pure cost/performance
Moreover, "closing the gap" could make them competitive in the foundry space, but it would leave them in a different position than they occupied historically. It was their technology _leadership_ that provided a competitive advantage, and it's why they are still so famously paranoid about their intellectual property.
> Moore's law wasn't okay then and suddenly just stopped in 2018 when Intel started falling behind
Of course not. This has been a gradual process.
Well that's a walk back. You said they were irrelevant to the rest of the market. Now it's a foundry customer couldn't pay any amount of money. Of course. (Well not strictly true, Intel did try to get into the foundry business and did get a customer or two, they just didn't do well https://semiwiki.com/semiconductor-manufacturers/intel/7912-...)
> Limiting physics. It is very hard to continue to make devices smaller at historical rates. Nobody is keeping up with the historical 2-year node cadence any longer. And the "nodes" now are marketing smoke and mirrors, anyway; they don't actually mean that device density has doubled.
I don't see how the argument has been made. Limiting physics was limiting in 2015 as well. Progress has still been made, and progress will almost certainly continue to be made for at least several more node, just at a slowing rate.
There is no before/after I've seen that makes a rational case for them previously being able to compete and now not able to. "Moore's law slowed therefore Intel can't get ahead of TSMC" isn't really arguing anything.
A crude approximation is money. Investment in R&D and fabs. So that's something you could argue, but when you look at Intel and TSMC, it's not obvious that TSMC has vastly more money to spend there.
> They cannot publicly pull the plug for two major reasons:
[...]
This is going into conspiracy theory territory. You think they're lying to investors and customers and the US government to fraudulently get subsidies investment and customers?
Look, you said they were going to not attempt to invest in regaining an advantage, and that they would focus on getting the most out of their trailing nodes. That's just not what's happening at the moment. And with the amount of money Intel is making still, I think they would be mad to give up now. I don't know what writing on the wall there is everybody thinks they can see here.
I was maybe not clear: Intel was irrelevant from the perspective of a foundry customer because their service was not for sale; it didn't matter what node they were at. Of course for foundries themselves, Intel was (and to some extent still is) relevant as a benchmark and as a technology leader.
> I don't see how the argument has been made. Limiting physics was limiting in 2015 as well. Progress has still been made, and progress will almost certainly continue to be made for at least several more node, just at a slowing rate.
I agree with this statement and it's compatible with what I said above.
> This is going into conspiracy theory territory.
Uh.
Do you think Google is being completely transparent with the public about things that are being said about Google Cloud in board meetings?
I didn't say they're going to just fall over dead. Obviously landing huge subsidies or government support would change the equation. But while Intel management is certainly aware of the state of things, they're not going to ensure a self-fulfilling prophecy by broadcasting pessimism.
What they're not going to try to do is regain the process crown because it's a fool's errand. That's what I said. It might bear emphasis that that is different from competing at the leading edge.
That doesn't follow. There is a big, big difference between not being completely transparent and making false or misleading statements to investors.
You said they've got no prospect of performance leadership and are going to focus on just getting the most out of their existing fabs. This is not a fact and not supported by their statements and actions so far. It's a wild guess and pretty unlikely to happen given their investment pipeline. Some time in the future, possibly. As a result of the circumstances right now when they are making money hand over fist with their leading edge nodes and building new fabs? No.
There are plenty of process improvements to be made at the trailing edge of development, where the majority of manufactured chips actually are. The leading, most expensive device nodes which are most clearly bumping into physical constraints, will gradually be less and less representative of the industry as a whole.
But the leading edge is far from standing still. I GAA and such pan out well that could be a very significant bump for the next few nodes. People act like everything stopped because "Moore's law is dead", but when you look at the rates of improvement in most other manufacturing sectors, semiconductors are still advancing at an enviable pace.
This is really not what I said anywhere. What I said is that the pace of advancement is slower now, and that process supremacy is no longer sufficient to guarantee performance supremacy.
That's a problem for Intel, since it was a core assumption of their business model.
FinFET and GAA do achieve slightly better electrostatic control which results in lower leakage and slightly faster switching speeds, but gate length doesn't decrease much anymore, which puts a ceiling on switching speed improvements (thus frequency - GHz), and gate capacitance (linked to power consumption).
Increasing densities mean more avenues to explore specialized circuitry and dark silicon. Basically, more transistor to play with.
This is running very contrary to the previous paradigm with Moore's Law, in which specialized circuitry and specialized accelerators were not too deeply investigated as software implementations would gain so much performance that they would easily surpass them, while not needing to be redesigned for a newer node.
I'm now hoping for (and expecting) a revolution in maskless lithography at some point. That would democratize ASIC design (at 14nm, current multi-project wafers prototyping costs are around 10k€/mm², largely due to mask manufacturing), and make it much easier to explore the third dimension (thus making everyone more competitive with last gen).
[1]: https://en.wikipedia.org/wiki/Dennard_scaling#Breakdown_of_D... (although that article is a bit terrible).
The technology has improved since then. And it will continue to improve.
Transistors are still at the hearts of ICs, so every bit of efficiency gained there has tremendous impacts on system-level performance and efficiency.
Some of the biggest possible gains on the horizon might come from circuit designs, notably adiabatic computing.
No, my intention was actually converse of that, saying that transistors still shink, even in leading edge nodes today, despite popular perception.
> I feel like I have to mention that transistors haven't been meaningfully shrunk for years, although there have been plenty of innovations regarding their layout, and photolithography processes have gained in precision.
TSMC 7nm->5nm increased transistor density by 80%. There were other changes to minimum cell size too, but both are 6T in this case so the overall device that can operate as a transistor in a VLSI does absolutely shrink.
Dennard scaling ending isn't about not shrinking, it's about power density going up. Even then things are not standing still, practical designs absolutely get denser and do more work even in the logic parts of the design. It's just doesn't double every couple of years anymore.