Intel Unleashed, Gelsinger on Intel, IDM 2.0
stratechery.com
stratechery.com
1. This is Intel's proverbial hailmary rather than "unleashed". The industry is moving on from x86 and with its other bets (e.g. radio modem, mobileeye) not paid/paying off, the revenue will quickly nosedive within the next few years.
2. IDM 2.0 is an involuntary move due to Intel's loss of both performance leadership (to AMD/ARM) and fabrication node leadership (to TSMC). Clearly their factories are becoming liabilities and opening up to outside business cushions the freefall.
3. 2023 is a long time away in the silicon industry. As pointed out in Stratechery, 7nm is actually further delayed than what Bob Swan previously announced. Apple is rumored to be on TSMC 3nm this year and AMD will likely move down to 5nm by next year. Unless Meteor Lake is a massive performance uplift, then it may not change the overall trajectory. Intel's saving grace at this point is the capacity constraints at both TSMC and Samsung.
No idea how to measure the impact, but I'm sure it's in the tens of billions.
It doesn't even matter if it's right or wrong - it matters because the headline of 7nm vs 5nm sucks for INTC stock.
I think Intel could easily put some marketing language about shifting to an "industry standard measurement practice" or something similar (remember, it's not for techies, but the people who buy INTC stock).
If everyone did, then things like GME would never happen.
Entirely precise yet only partially meaningful, which is just as true of the chase for node bragging rights.
So there's some karmic justice in Intel being pilloried for being behind in this way.
Separately, what you are saying is true -- size is not the be-all-end-all of processor performance. But GP is referring to the dislocation between nomenclature and geometry.
Interestingly, the shift to multi-core on the desktop hasn't been the big problem that a lot of people feared it would be. I remember hobbling on crutches cross-country to SF for an Intel quad core (?) launch and there was a lot of hand-wringing on the topic but mostly things worked out.
This bit Apple too when they boasted of their PPC clock speed, only to fall behind in both clock speed and performance. (I'm a Mac user: this is bipartisan criticism).
I'd say I'm glad those days are over, but of course the deceptive metrics have merely been replaced with different deceptive metrics.
Not sure where that rumour are from. But TSMC 3nm is next year.
If Gelsinger didn't lie, and the original 7nm schedule and spec were unchanged, the 100% increase in EUV usage would actually put it close to if not on par with TSMC 3nm. Although they will still be a year and a half behind TSMC.
And technically 7nm are taping out this year. The only reason why volume product wont come until 2023 is because Intel has a SuperComputer contract to fill with the US DOE.
Both Intel 5nm and TSMC 2nm will transition to GAAFET. But I guess Intel ( Pat ) want us to focus on the strategic changes rather than the detail of future roadmap.
It depends on how Apple maneuvers. Mobile chips (A15) are high volume so very unlikely to get moved but it's certainly possible for the lower-volume Mac products, e.g. M2/M1X for Macbook/iMac, to switch onto leading-node 3nm if yields are good.
Apple is certainly incentivized as it'd allow them to continue the narrative that their own SOCs are superior to anything else on the market.
Hasn't this narrative been largely shaped around performance impact on customer experiences?
Recall the announcement video which was decried as "marketing hype" by some tech voices prior to release.
If so, M2/M1X should be able to continue this narrative on a performance basis alone regardless of whether they make the jump to 3nm.
A big chunk of Apple's power advantage (which can translate to better perf as well, since the key limits are thermal) can be attributed to the process node, though we won't know how much for sure until we get an apples to apples comparison. I agree with Gap, retaining the process node advantage will help sustain and entrench the perception that Apple's chips are better, easing the transition.
7nm risk production: April, 2017
7nm first Apple SoC: Sept, 2018 (A12)
5nm risk production: April, 2019
5nm first Apple SoC: Sept, 2020 (A14)
3nm risk production: 2021Q4
Following this trend, when do you expect the Apple 3nm SoC to debut?
Risk production is when you start testing wafers and design for performance, yield, other tuning and bugs. Those are rarely if ever sold on the market ( Unless they are called CannonLake ). So realistically it is still 2H22 for 3nm.
And generally speaking leading edge node rarely if ever go to large die size SKUs first ( That is MBP and iMac ) simply due to yield and cost. ( Unless your customer is US DOE buying chips for their SuperComputer called Aurora. )
Intel can't even solve their 10nm yield problem. Its great to have another foundry but Intel needs to step up.
PREVIOUS (incorrect): You can buy them (the "Rocket Lake S" platform) beginning March 30th, I believe. One issue they are having with 10nm (though not intently a fatal one) is that it doesn't seem to achieve market-leading single-thread performance without being ridiculously power-inefficient. In lower-power environments, e.g. in laptops, the 10nm chips seem to be "fine", in that they are improvements over 14nm+++. They're still not market-leading, but they are an improvement and not at the cost of power (relative to previous generation).
Uh, what? Mobileye bagged $333mm in revenue last quarter.
See for yourself:
https://d1io3yog0oux5.cloudfront.net/_9ea32bedbda3919ea6d9a1...
"Hey we have a world class [whatever] over here. Wonder if we can sell it to the world as well as using it for our business?"
Which is what Intels move looks like to me. Go hard on Fabs, and sell production to others as well.
Cortex series are pretty potent for commodity chips. Now Apple does have their completely different cores that just happen to use ARM instruction set, but I think most fabless are not focused on the cores themselves.
I should have said arm64 chip. AMD is fabless and they do spend a lot of effort on core design.
This is pretty bad for Taiwan geopolitically though... once US chipmakers have a plausible local manufacturing alternative, the US is less likely to risk WW3 by standing up to China if/when it tries to annex Taiwan.
Then as the mainland market developed they increased their mainland footprint as a hedge (makes it harder for the Chinese government to attack TSMC, though as we've seen from Alibaba, Xi appears happy to "move fast and break things"). That did in fact forestall investment in a competitor until recently.
They have also been expanding outside China-Taiwan, distributing their capacity and developing support in the USA and Europe. Worst case (actual shooting war between China and Taiwan) they would have a lot to fall back on.
Hon Hai has followed the same geostrategy (to even greater extreme) with its Foxconn subsidiary.
Very unlikely. In case of a shooting war, the industry will run out of consumables sooner, or later. As a fact, 90%+ of them come from the Taiwan/Japan/Korea triangle, and just any military move in the region will be effectively shutting the industry down globally.
And seeing men thinking so again, 100 years after, only tells that Raymond Poincaré was right.
TSMC's GIGAFABs (> 100K/month) are used for mass manufacturing bleeding edge consumer electronics. They are still build only in Taiwan.
[1] https://technosports.co.in/2021/03/03/tsmc-to-build-5nm-plan...
Biden has to repair our US hegemony. Trump sure hurt it a lot in spite of his anti-china rehtoric.
Selling weapons to taiwan doesn't contain china. We will send our own guns if a hot war is brewing there.
Taiwan recognition is not the same thing as meaningfully containing china. The TPP and asia pivot from obama was the last time that america seriously tried to restrain the rise of china.
Trade deals are ugly.
But if your geopolitical goal is to prevent Asia from moving closer to China, you can, and we have, done worse than the TPP.
And people still celebrate Trump for being tough on China...
All of these states need to be strong enough to be a credible deterrent to invasion on their own.
US support is neither speedy nor certain. And a speedy capitulation by Taiwan compromises the US's ability to help even if willing.
US does not have expertise to make modern non-nuclear subs. China can intimidate anyone in Europe not to sell. It was fortunate for Taiwan that Chinese pissed off Japan enough. Half a dozen of Soryu knockoffs can shut down PRC maritime trade if needed.
This isn't an attempt to start a debate, just saying that the U.S. won't intervene unless there's material gain, regardless of what the voters want.
It's really more of a petrol bomb, yeah.
Realpolitik is what it is, but in this case standing up for the cause of Taiwanese freedom against (mainland) Chinese aggression aligns fairly well with the current political climate in the US. I'm willing to accidentally do the right thing if we have to.
So long as TSMC and Samsung remain competitive (note, this doesn't mean better, just close enough -- better is also fine), I expect the vast majority of AMD and Nvidia chips to be manufactured by those two.
Why not? Nvidia seems to be very pragmatic in their foundry picks on what's available for example.
Yes and no. This is also a possible signal that Intel is getting ready to go fabless. AMD made similar moves with Globalfoundries before parting ways.
Intel may make it modular, but Intel will still own those fabs. I do not see them spinning fabs as separate business at least for the decade of 20s.
I also think China/TSMC is more of an economic than an existential risk. The US military doesn't need 2nm semiconductors. US manufacturers are more than capable of supplying essential US supply chains. They're just not cutting edge like the East Asian fabs are.
I don't know why you are surprised as lots of HN users keep saying that we either already have FSD as promised or that it will be here soon (in the Tesla beta?). People believe the most crazy things, especially if they have already bought into the ecosystem somehow (see also Apple Vs. Android discussions).
I doubt it changes much to the demise of modern world. The industry will still go down given that you have many, many, many more things in the semiconductor supply chain that is run by some single vendor in the world from Taiwan, and that includes consumables too.
US may get a modern fab, or two, but it will still eventually run out of many know-how intensive inputs for those fabs.
If Intel can pull this off it will be one of the most impressive recoveries in history, up there with Watson at IBM, Gerstner at IBM, and Jobs' return to Apple.
And if Intel can pull this off I wonder if someone will be able to do the same for GE.
Compared to GE... there's been more structural changes to GE over the years which would hurt its ability to regain a useful position in vertical integration from where they're at - they need multiple decades of long term planning to get around this, especially as their industrial side becomes increasingly costly. This is unlike Intel, who is still in a "good" position to pivot on the reality that x86 is a lot less valuable compared to 2009 into a more up-to-date vertical monopoly.
Gerstner was astonishing too. I just thought of him as the "cookie guy" when he showed up. But he saved a company that was spiraling towards the ground...he didn't restore it to its former glory but did bring it back to a successful life. An achievement subsequently squandered.
https://spectrum.ieee.org/tech-history/silicon-revolution/bu...
S/360 was/is a big deal in history.
Additionally, US politics is now aligning across the aisle against China, and relying too heavily on TSMC looks like a potential future geopolitical problem. Intel is the only firm that can realistically keep up or out-fab TSMC and thus the USG will be loathe to let it even look like it might fail. Simply having fabs physically within your territory but managed by a remote firm, is clearly no substitute for having fabs actually managed by a domestic company (consider how many ways there must be to do the equivalent of SSH-ing into a fab, or otherwise subtly sabotage it from HQ).
Samsung?
Realistically, Intel needs a route to actually playing at TSMC's game, currently they are instead working on paying TSMC to make them chips.
Also, nitpick here, but the TSMC problem is that China might roll in tanks, not that China controls it. Realistically speaking, it's a reasonable cause for US millitary involvement if China did. Lol, there's my World War III Shark-jump Theory: TSMC.
In the end getting a CPU is normally more important then getting the best CPU.
Personally, if Xe production funds TSMC's 3nm node that will be bought out by Apple and AMD then Intel probably should realize that the limited money they will make from GPUs will be dwarfed by the losses from improving their competition.
TSMC have fab in China and currently building one in the US. TSMC is also opening an expansion in Japan. https://www.reuters.com/article/us-taiwan-tsmc/tsmc-to-raise...
the latest tech is probably going to stay in Taiwan but judging from all the recent moves by TSMC. it look like they are spreading their wings to reduce the risk.
Lithography is definitely key, but all the equipment and process must work together.
I don't think enough people fully appreciate that just about the entire world's leading edge supply is located within cruise missile range of China. The CCP could cripple the global economy in under an hour. Having supply chains out of their reach is not a nice to have, it is a must.
From what (little) I know of the industry this isn't a good sign. Engineers working on the fab processes and packaging have to work closely with designers throughout the product design cycle, repeatedly taping out variations to fit onto the process as it is built out.
That's part of why Intel historically used the tick/tock design cycle, designing a new thing to be built on a new process has a multiplicative effect on the time to market because of the way the unknown-unknowns of both sides impact each other, and the large amount of specialized labor to mitigate everything.
It doesn't help that they lost a ton of institutional knowledge over the last decade to their competitors through layoffs disproportionately hitting senior staff.
Conversely, if he were to say "it's worse than I feared," he won't personally be blamed because the problem preceded his return to Intel.
So I think the claim can be viewed with a greater sense of honesty than before.
Note that Intel's current fab problems were well underway when there was an engineer on top, and that engineer was a fab person. Pat is not.
I'm under the impression 10nm is largely working by now, it just took a long time to become competitive with Intel's ultra optimized 14nm process. Supposedly they have already shipped 100K+ units of Ice Lake Xeon Scalable CPUs: https://www.anandtech.com/show/16539/intel-ice-lake-xeon-sca...
Like, say, giving the engineers more say in the goals for the next process(es) to make them achievable. Talking, not ruling.
nice to see someone addressing that. I welcome technical leadership back to intel.
Intel foundry have to be a pure play in order to win customers trust. Apple move away from Samsung and now a core customer of TSMC, trust is an issue.
This. It is the reason why Apple dont Fab their SoC with Samsung Foundry.
If you have no direct competition with them then of course you can Fab with Samsung.
Samsung talked about Firewall between Samsung Electronics and Foundry, but knowing the Korean culture and Samsung, that wouldn't be much of a good idea. I would even bet Foxconn CEO Terry Gou had a few word of advance to his friend Tim Cook.
But given the combination of losing the node advantage to TSMC and performance advantage to AMD & ARM, I'm guessing Intel has accumulated inventory and this poses a business risk. Opening up their fabs in the future is a way to mitigate this risk; essentially becoming both a chip seller and external manufacturer.
The other reason is probably textbook monopolistic behavior: vertical integration means if you want top tier chips, there was no other option but Intel. Often overlooked is how closely correlated fabrication node is to chip performance, e.g. if Apple's M1 is fabbed at TSMC 7nm instead of 5nm, the performance wouldn't be so widely lauded.
When fab is 4-5 years old, they start to manufacture other products. Different CPU's, chipsets, SoCs, microcontrollers, NICs, Wi-Fi and Bluetooth chipsets, vechicle SoC's for Intel and Intel customers.
This is the microeconomic equivalent of the "dutch disease"
Intel becoming the best foundry in the world might be the only way they can sell vastly more chips, but it's not easy; and it's not being headed by a turnaround king, but rather the master of harvesting (why is there still vmware around since virtualization became a feature of chips and operating systems?)
I'd have to say - if you think VMware is delivering only virtualization to their customers then you probably haven't looked at their portfolio as of recent.
They have over 60 unique products that target many different areas of compute in the enterprise. Gelsinger is generally attributed for keeping VMware relevant. Not everybody is a consumer of a hyperscaler and some customers want extended enterprise functionality for virtualized (VM and containers) environments.
Because VMware (mostly through acquisitions) makes a whole bunch of other stuff now too.
Because, besides the need for a frontend, there's a lot of devices to emulate.
If you take a look at https://www.vmware.com/products.html you'll note the majority of VMware's products are NOT related to compute virtualization (and a very significant entry from that category, vSphere Hypervisor, is given away for free).
> We are committed to ensuring this capacity will support commercial customers, as well as address unique government and security requirements in the U.S.
Geographic distribution makes a certain amount of business sense and getting those nice DoD contracts makes even more.
What doesn't make sense is China having Wing Attack Plan R ready to go against TSMC and Samsung.
IDM in most people's minds is Intelligent Dance Music.
But, in Tech I am familiar with many people using IDM for IDentity Management. When I was associated with a different Tech vertical than I am now, people used IDM as an acronym for Integrated Document Mangement.
IDM = Industrial Dance Music for me. I guess there are a few definitions floating around.