The Fall of Intel
- Thread starter Eric
- Start date
Another thing I think a lot of “government subsidies bad” people miss about Intel is that this strategic and value add capability is good in the long run for the US due to our design dominance: even as Intel fades, we do dominate high value add compute which relies on leading fabrication as the major choke point.
Sure, Taiwan being blown to bits would hurt China too but less than us on net, which would also create openings for inferior substitute goods using an older SMIC node in the event of a war’s aftermath. Or in a blockade event, if they take the fabs, even though in principle TSMC would rely on goods that are sanctioned by US against the PRC — most likely if the fabs were intact there would be immense pressure to remove those by US allies, and Chinese firms would then have access to leading processes (and also eventually, I am sure, flip the switch on us if they can ever develop EUV on their own).
It’s just way too important that we have good fabs that are also technically and reasonably economically competitive.
Much more important than like, domestic steel which we can buy from Germany or low grade from Pakistan or Turkey.
Sure, Taiwan being blown to bits would hurt China too but less than us on net, which would also create openings for inferior substitute goods using an older SMIC node in the event of a war’s aftermath. Or in a blockade event, if they take the fabs, even though in principle TSMC would rely on goods that are sanctioned by US against the PRC — most likely if the fabs were intact there would be immense pressure to remove those by US allies, and Chinese firms would then have access to leading processes (and also eventually, I am sure, flip the switch on us if they can ever develop EUV on their own).
It’s just way too important that we have good fabs that are also technically and reasonably economically competitive.
Much more important than like, domestic steel which we can buy from Germany or low grade from Pakistan or Turkey.
Based on the past decade of intel lies vs. results, I'd wait until I see actual production silicon before believing anything they say
Apple built a whole suite of tiny form factor machines no doubt based on intel performance projections, and look what happened there...
- Joined
- Sep 26, 2021
- Posts
- 6,318
- Main Camera
- Sony
No way. Martial law was just declared in Korea.
Not to mention Samsung’s history of grift.
Yoused
up
- Joined
- Aug 14, 2020
- Posts
- 6,840
- Solutions
- 1
nVidia already has a large SoC/chipset called Grace-Hopper which has ARM processor cores. They have an ARM license, and even tried to buy ARM out a few years back. That they would want to own x86 would be hard to understand, unless the goal is to profit off dwindling sales as they slowly strangle it.
That was my idea. Not sticking with x86 but offering a performant migration path for it.
- Joined
- Sep 26, 2021
- Posts
- 6,318
- Main Camera
- Sony
nobody’s stopping them. Apple did it without an x86 license.
Fair enough, but safer than anything from China or TW. That said, I obviously prefer keeping IFS’ ownership domestic and the future of it pure play, with a fat stack of cash from USG contingent on future rollouts (also, they need serious permitting clearance, our regulatory policies are a disaster for fabs and manufacturing).
But if I had to choose to please people that hate subsidies and gov intervention* who I’d rather have take merger than shuttle Intel entirely.
Again, not my ideal. I’m in favor of throwing more money at Intel (the fabs) strapped to a contract, spinoff & replacement of leadership, on the order of 100-200B over the next decade. It’s very much doable for us to have this, purely a question of will and in the grand scheme the cost is quite low IMHO.
Also probably wise to partner (not symmetrically re: control of the IP etc but) with the Europeans and possibly Japanese on this. Economies of scale are important as is some cost reduction from outsourcing or buy in from allies with a similar interest in friendshoring and risk pooling away from TW & China or in the more aligned EU case, away from East Asia.
Yeah it makes literally no sense for Nvidia to go adopt X86 or buy a license. They have much more control with an Arm license and I suspect even Qualcomm with the recent legal battle isn’t upset enough to sincerely believe Intel & AMD are more trustworthy.
Also to the extent ISA reduces some design barriers and overhead I imagine they’re quite fine leaving it behind.
The migration path is just adopting Arm & emulating crud well enough. And porting.
That’s it. No “what if we had an X86 core or two for emulation” Frankenstein proposals or wacky Transmeta stuff, just basic PRISM with AVX2 & relatively basic (impressive yes but I mean, big scheme) TSO support in Arm hardware is enough for reasonable emulation going forward especially with Microsoft continuing to support that pathway.
Microsoft also guaranteed doesn’t want Nvidia licensing from Intel and AMD with X86. For one it’s only one other form and MS wants a bit more openness than that, I strongly suspect internally they’d like to just see everyone swap over eventually. Even if a Nvidia-X86 gambit succeeded it’s volatile and who knows what would become of it, Intel/AMD might just use them to fend off Arm and terminate the license eventually. Nvidia certainly would know this too. Arm even with Cortex (which is a conflict of interest yes) still has incentives more aligned with licensing an ISA than Big Red & Blue.
That’s it. No “what if we had an X86 core or two for emulation” Frankenstein proposals or wacky Transmeta stuff, just basic PRISM with AVX2 & relatively basic (impressive yes but I mean, big scheme) TSO support in Arm hardware is enough for reasonable emulation going forward especially with Microsoft continuing to support that pathway.
Microsoft also guaranteed doesn’t want Nvidia licensing from Intel and AMD with X86. For one it’s only one other form and MS wants a bit more openness than that, I strongly suspect internally they’d like to just see everyone swap over eventually. Even if a Nvidia-X86 gambit succeeded it’s volatile and who knows what would become of it, Intel/AMD might just use them to fend off Arm and terminate the license eventually. Nvidia certainly would know this too. Arm even with Cortex (which is a conflict of interest yes) still has incentives more aligned with licensing an ISA than Big Red & Blue.
It’s not a hypothetical, we know Intel 4 to TSMC N3B resulted in about + 12-18% performance iso-power improvements or about 25-35% iso-performance power reductions with* Intel’s new Lion Cove architecture on N3B vs the last one on Intel 4. That Lion Cove has major changes towards at least somewhat improving efficiency by going wider (iso-performance, lower the voltage and go slower) and keeping data closer to the core with a big L1/0. Is said core actually toast good? No lol but that’s not the point, it’s better than what they had.Based on the past decade of intel lies vs. results, I'd wait until I see actual production silicon before believing anything they say
Apple built a whole suite of tiny form factor machines no doubt based on intel performance projections, and look what happened there...
So first of all that isn’t iso-architecture, which puts a hard bound on how much of this is a TSMC improvement from an electrical perspective, even if nothing improved architecturally (doubt). Now if that were true it still just implies they got about one normal vanilla to vanilla process node gain in power/performance going from Intel 4 to TSMC N3[B though Arrow Lake might be N3E iirc].
Now sure I do suspect most of the gain is TSMC, but even if it’s like 65-80%, still not bad either way.
That’s… good because it shows how Intel’s last FinFET fabs really aren’t that terrible from a performance and active power perspective, but what is terrible is their architecture. If Intel PC guys were right about how good Intel is then a chip [Lunar Lake] with new fancy PMICs and power delivery, tons of cache and an SLC, all on N3 would have narrowed the gap with Apple, at least an M2. It didn’t. They are tied with Qualcomm’s first screwup Junior Varsity core on N4P.
Intel 4 -> Intel 3 (last FinFET nodes) -> 20A (scratched now) -> 18A.
Intel 3 is being used for a Meteor Lake tweak and refresh fwiw and from the early clocks at the same TDPs it seems like they probably weren’t lying that Intel 4 to Intel 3 is a significant power/performance boost.
x.com
Why didn’t they use that for Lunar Lake etc? Well probably it wasn’t ready in time and is still lower volume which counts, and they signed a bunch of wafer contracts with TSMC back in 2021 for N3 specifically to get around this exact uncertainty, probably wisely.
But the point is there is some reason to believe Intel isn’t totally lost on some basic parameters of importance. I trust that they have more potential than Samsung anyway.
Whether or not USG will be as protectionist and active as Seoul will for Samsung, I am pessimistic on sadly. But still.
Intel 3 is being used for a Meteor Lake tweak and refresh fwiw and from the early clocks at the same TDPs it seems like they probably weren’t lying that Intel 4 to Intel 3 is a significant power/performance boost.
x.com
x.com
Why didn’t they use that for Lunar Lake etc? Well probably it wasn’t ready in time and is still lower volume which counts, and they signed a bunch of wafer contracts with TSMC back in 2021 for N3 specifically to get around this exact uncertainty, probably wisely.
But the point is there is some reason to believe Intel isn’t totally lost on some basic parameters of importance. I trust that they have more potential than Samsung anyway.
Whether or not USG will be as protectionist and active as Seoul will for Samsung, I am pessimistic on sadly. But still.
- Joined
- Sep 26, 2021
- Posts
- 6,318
- Main Camera
- Sony
heh heh. Transmeta.
If anything Korea handled a deranged unpopular president's attempt to stage a coup much better than expected. The president was elected in a large part by being an anti-feminist and a supporter or mens rights while promising stronger legal punishments on criminals, more communist bashing, etc, and he has proven to be absolutely incompetent even in this coup attempt. Stop me if you've heard this all before somewhere else.
Speaking of undeserved leaders, I remember thinking the initial cheer for Gelsinger's hiring seemed off, since it felt like a blind faith of the CEO being an engineer rather than a business guy, that came across as yet another blind STEM faith. Either way Cliff's call on him turned out to be correct.
I often think back to Otellini's claim that Intel could've been the chip supplier for the first iPhone if they were willing to sacrifice the margin. Would that have been an ARM designed and manufactured by Intel? Could that have prevented the amazing rise of TSMC given there must've been a ton of Capex enabled by Apple's success for the cutting edge processes?
- Joined
- Sep 26, 2021
- Posts
- 6,318
- Main Camera
- Sony
Welcome!
FWIW, I think a technical CEO is the right call for Intel at this juncture - I just knew Gelsinger wasn’t the right guy. You need an outsider who can be honest with him/herself about what really needs to be done. That was also the case when Gelsinger was installed - unfortunately, the situation now is in some ways worse than it was then.
Remember that Intel had DEC’s StrongARM for a short while. In an alternate history, they would have nurtured that team (I knew some of them and they were very good), kept the radio team, and not been stuck at 14nm for year after year, and then they’d be Qualcomm, but with the best fabs in the world. Apple would never have had to start its own chip skunkworks, and Intel would own mobile. They wouldn’t have had to panic and hire Gelsinger, and their CEO might have actually done something about AI before it was too late. But nope.
Intel 18A node yields reportedly at a dismal 10%
A South Korean media outlet reports Intel's 18A yields are worse than that of Samsung Foundry's second-gen 3 nm node. Apparently, they're at 10%, effectively rendering the node not viable for mass production.
www.notebookcheck.net
Yields of 18A are reportedly a “dismal 10%”. As the article states, 18A isn’t yet supposed to be available for mass production so yields have time to improve. However, a lot is riding on Intel’s 18A node and in their race to surpass TSMC, Intel, like Samsung, may have tried to stuff too much too fast into the final node of their fab death march. If this is accurate, problems in 18A could be an even bigger reason why Gelsinger was forced out as, even more than the AI failures and Raptor Lake woes, he bet the company on this 5 nodes in 4 years death march succeeding.
Of course TSMC’s GAA node is still in development. While so far their more cautious approach appears to be vindicated, they may struggle with yields as well. We just don’t know yet.
Yeah. Classic case of missing the plot IMO.
Yep. There’s a real timeline where Intel takes some kind of new core design and ISA seriously designed for low power with no exception (down to the fabrics which, theirs are abysmal and it’s part of their problem) and wireless, seriously.
They’d have then been able to amortize fixed costs for CPUs across everything from a smartwatch chip and phone or tablet to laptops, DIY and servers, embedded. Would be Qualcomm 2024 on steroids but their monomaniacal fear of undermining their monopolistic, lazy cash cow and lazy vertical integration gains drove them to ignore this.
The same thing is really true of the vertical integration and fabs. With the fixed costs of fabs rising since 00’s, the writing was on the wall by 22/14nm in the early and mid 10’s, roughly. You need volumes to drive that economically and reinvest, and mobile was actually a key driver for TSMC and Samsung. I think even if Intel had done everything right, on some level they’d want to be outsourcing. If Intel were Apple’s supplier and they had never gone vertical for instance you’d probably still end up having some other competition via Mtek and Qcom themselves etc, though the vertical integration [suggesting the fabs are closed off] would be vastly more tenable if Intel had mobile volumes and weren’t on the cusp of losing laptop and server volumes of course.
I’m skeptical of this. Panther Lake is going to be on 18A and while die dize is related to yield and node differences there are less obvious than with larger dice, this is missing something. Panther Lake is less tiled than Meteor Lake — it should be a CPU + SoC tile on 18A and GPU on N3 — and that 18A tile won’t be too far off from the dice used in a phone either + or -.Intel 18A node yields reportedly at a dismal 10%
A South Korean media outlet reports Intel's 18A yields are worse than that of Samsung Foundry's second-gen 3 nm node. Apparently, they're at 10%, effectively rendering the node not viable for mass production.
Yields of 18A are reportedly a “dismal 10%”. As the article states, 18A isn’t yet supposed to be available for mass production so yields have time to improve. However, a lot is riding on Intel’s 18A node and in their race to surpass TSMC, Intel, like Samsung, may have tried to stuff too much too fast into the final node of their fab death march. If this is accurate, problems in 18A could be an even bigger reason why Gelsinger was forced out as, even more than the AI failures and Raptor Lake woes, he bet the company on this 5 nodes in 4 years death march succeeding.
Of course TSMC’s GAA node is still in development. While so far their more cautious approach appears to be vindicated, they may struggle with yields as well. We just don’t know yet.
Of course it could be another 10nm I just wouldn’t take this too seriously. No clients that have *officially* announced their use of 18A rather than interest have backed out. What we do know is the HD cells aren’t good enough yet on density and leakage, which is why 18AP is coming to amend that with Intel explicitly noting it for mobile, and this is also the reason Qualcomm apparently dropped early interest.
Simplified $20,000 wafer cost adjusted for various dice. In practice it might be better than this if your marginal area is a lot of parallel logic with redundancy (e.g. SoC blowing up from 150mm^2 to 250mm^2 where the main difference is GPU and encoding size) but still.
Similar threads
