Intel Core Series 3 will be produced in the US, reducing dependence on TSMC
Posted by sr_local@reddit | hardware | View on Reddit | 81 comments
Posted by sr_local@reddit | hardware | View on Reddit | 81 comments
frogchris@reddit
Wow like Intel has been doing for the past 20 years. Amazing lol
DehydratedButTired@reddit
They haven’t been doing it though. They have been struggling to keep up and kept using a worse process because they couldn’t make a similar process work. They only moved to TSMC because they had to.
Dstln@reddit
They're literally pumping out chips all day right here in Oregon. These factories aren't running for show.
DehydratedButTired@reddit
So is global foundry and SMIC but they aren’t making cutting edge chips like TSMC and Samsung. Intel is still playing catchup. Intel’s Oregon facilities make 18A and 14a nodes but what they are producing there is not cutting edge. These areold processes refined to a high level. There is research going on there and elsewhere. We will see where their fab ends up.
tacticalangus@reddit
18A and 14A are absolutely leading edge process nodes.
Why do you claim they are "not cutting edge"?
nanonan@reddit
If it's behind the actual cutting edge, it's not cutting edge. They are closing in, but they aren't there yet.
tacticalangus@reddit
It is the first process node with both BSPD and gate all around with actual products on the market that you can buy right now. In terms of PPA it is perhaps on par with N3P. N2 has higher PPA than 18A but most products with it haven't launched yet and the differences in PPA between these nodes are getting smaller and smaller.
It takes some serious mental gymnastics to claim 18A and especially 14A aren't leading edge or cutting edge.
raulgzz@reddit
They powerful cores are way behind Qualcomm and apple's powerful cores, their good igpu Arc B390 is still made by tsmc.
Intel's 18a is at M3 level at best but it uses a lot of more power than tsmc-apple M3.
tacticalangus@reddit
You're comparing an Intel process node to an Apple SOC made on a TSMC N3 process node. Intel's P cores are quite behind compared to Apple, it isn't 18A that is to blame for that difference.
Lunar Lake on TSMC N3B has worse performance per watt compared to Panther Lake on 18A using very similar P cores.
B390 is a good GPU but it is mostly good because it has a solid microarchitecture in Xe3. The 4 EU and 2 EU versions of that are made on Intel 3 and 18A. The Intel 3 version still demonstrates competitive performance per watt.
The take away is that Intel designed P cores are behind and need to improve significantly. Intel has a good GPU architecture. 18A and TSMC N3 are similar in PPA.
jmlinden7@reddit
18a and 14a are literally cutting edge. They might be a few percent worse than TSMCs newest nodes but they are literally still on the cutting edge
Exist50@reddit
14A doesn't exist yet. 18A is essentially aligned one node behind TSMC. What does cutting edge mean if not the most advanced?
Johnny_Oro@reddit
Even though 18A is like N3 level, Intel got to BSPDN before TSMC. It's still leading edge in some ways.
EnglishBrekkie_1604@reddit
I think calling something that’s a gen behind cutting edge is still fair, especially considering that the death of SRAM scaling has absolutely murdered node gains.
Before the EUV era Intel TSMC and Samsung were all at the cutting edge, but Intel also had a clear lead.
Into the EUV era TSMC had both a clear lead and the only nodes actually worth using, with Intel and Samsung both trapped in purgatory with nodes that were only attractive on paper.
Now we seem to be coming back to all 3 being viable, though TSMC still has an obvious lead.
But I think the N3 saga shows that TSMC isn’t invincible, all it might take for them to lose their lead is them making the wrong bet and someone else making the right one, and ironically enough I think Intel would be the most likely one to win in that scenario given their historical inclination to saying “fuck it, we ball” and coming out on top.
Exist50@reddit
I mean, doesn't that pretty much directly contradict the definition of the term?
Is that not where we're still at today? Intel can't get a single major customer, and Samsung's bled most of theirs to TSMC, at least for leading edge.
Uhh, they've spent the last decade failing at both their radical bets and their not so radical ones. They need to walk before they can run.
EnglishBrekkie_1604@reddit
It’s semantics IMO. Personally I draw the line at “is this in a performance tier good enough to be used for a non-premium but current gen mobile chip”. I can definitely see the argument for just whoever’s in first is the cutting edge, but I don’t think TSMC’s lead is quite big enough to totally write off the alternatives now.
I’d say this was true a year or 2 ago, but I think they’re out of the pits now. 2 years ago we had Intel 4 and Samsung 3nm. Those are 2 truly, truly worthless nodes. Now Intel with 18A is mostly competitive with TSMC, obviously behind N2 but if you want something “good enough” that isn’t TSMC then it should be competitive, but it’s pretty held back by lacklustre design tools. Samsung SF2 is in the same boat but also comes with non shit tools and is thus attracting some real attention.
A fair cop. Honestly my take here is mostly fuelled by what I’ve read about 14A’s development, that is to say an order of magnitude better than 18A’s. IMO it’s interesting that the only rumours for Intel that have been fairly consistently positive recently are for Nova Lake and 14A; with everything else it’s eternal anarchy and doom.
Exist50@reddit
Intel 3 was about a node behind TSMC at the time as well. I'm not sure I see what the difference is supposed to be with 18A. Similar deal with Samsung. And in both cases, it's reflected in customer adoption.
What have you read? Because the exact same claims were made of Intel 4/3 and 18A as well. Remember when they cancelled 20A and tried claiming it was because 18A was doing so well?
EnglishBrekkie_1604@reddit
Intel 3’s an interesting one; it’s actually an alright node, but the biggest thing holding it back is its high cost of production, already saturated production lines (gets used for Xeon and they literally can’t get enough of it right now) and, again, PDK and EDA tools. Even if the node itself is workable, nobody wants to go through the asinine process that Intel designers are used to.
My personal bet is that long term, once they get better design tools, improve yields, and work out some of the kinks, 18A and 18A-P will get some orders long term. N3 isn’t getting less busy anytime soon.
I’m gonna be upfront and say I don’t have any inside connections nor the experience that you do, so take everything I say with several grains of salt.
Anyways the reason I think it’s going better is because almost all of the best news about 18A, 20A etc was from Intel themselves, like the legendary “Qualcomm wants to use us!” before Qualcomm proceeded to not use them. This was undoubtedly a bit of Pat coming through, that man seemed to have a chronic inability to say something wasn’t going incredibly. If there’s one thing Lip-Bu has unquestionably done right, it’s toning down the gloating.
With 14A though, a lot of the whispers have come from outside Intel; some of the people with industry connections I remember who down-talked 18A saying “it’s going badly, they’re bluffing” are saying “it’s actually going quite well this time, there’s some meaningful interest externally”. The fact I read that quite a while before the CTO (? may have been someone else) said something pretty similar made my ears perk up a bit. Of course, it’s Intel, for all we know they might set it all on fire, but it definitely seems less chaotic overall this time.
Exist50@reddit
Who specifically?
And I'll certainly give Lip Bu credit for being more tight lipped, but I don't see him hyping it up much either. And it's seemingly already slipped to ~'28. Though perhaps that's less of a delay and rather just what the reality was known to be, and Intel was just lying about it before.
nanonan@reddit
They literally do R&D in Oregon, not mass manufacturing.
EJ19876@reddit
The building towards the northeastern side of the RA campus in Oregon is a volume fab. It used to be called Fab 20.
Dstln@reddit
Wrong, they do both and plenty of manufacturing.
frogchris@reddit
No they used part of tsmc process. And stich them together though silicon via and chiplets. The base was always using Intel. They used tsmc because it as better, and having your design being limited to your foundry limited the competitiveness of your product. Intel silicon should always be using what is best and make sense instead of losing market share.
Intel was never going to keep using tsmc. Hence the billions of dollars in capex spending they did in 2020 and 2021. Pat Gelsinger did an all in gamble on the foundry service and it worked. The Harvard liberal arts mbas don't understand engineering and the time it takes to get results from large capex investments.
DehydratedButTired@reddit
Intel only had to use TSMC because they fucked up. What I said hasn’t changed. They couldn’t make their process work for a decade and a half, lied about it the whole time and then had to use their competitor to bridge the gap. Gelsinger for left holding the bag for sure and his gamble hasn’t paid off yet. They still need to never in their research and prioritize their fab development. Don’t get caught up in their hype and marketing. The proof will be when they can bring it all back and keep up with TSMC. There is no reason the should have fallen behind in the first place.
EnglishBrekkie_1604@reddit
I’m not sure Intel will ever be able to bring everything back in house, due to the simple fact that they don’t have a single node suitable for simpler tiles like the platform controller. Intel 3 is far more useful for other things because of how advanced it is and supply constrained it remains, Intel 4 is godless, I don’t need to explain why Intel 7 won’t work, and 14nm is ancient. In the future they might go to Intel 3, but something like N4 will perform the same but be cheaper, so they’re just as likely to keep it external for the sake of margins.
Plank_With_A_Nail_In@reddit
This is saying the exact same thing but using more words, well done I guess.
Johnny_Oro@reddit
Not quite entirely, it's got a TSMC N6 PCH die, although that's a cheap old node that makes up a small portion of the chip.
Ok-Parfait-9856@reddit
Isn’t the pch on the motherboard, not packaged with the cpu? I know Intel provides it either way
Straight_Loan8271@reddit
Not in laptops, usually. Putting the PCH on package with the CPU helps with thermal management and reduces power usage. The exceptions are the HX suffix mobile CPUs using the same die as the desktop chips
Ok-Parfait-9856@reddit
Okay, so for desktops the pch is on the motherboard right? Because I said that elsewhere and it got downvoted, but I’m pretty sure current desktop CPUs don’t have a pch on package. It makes sense laptops would package it for power management reasons. Thank you btw
Straight_Loan8271@reddit
Yeah. PCH is the Intel term for what everyone else calls a motherboard chipset. Which is a term that I hope anyone who's ever built a DIY system has encountered, because what chipset is on a given motherboard is the main thing that sets it apart in terms of features from other boards for the same socket
goldcakes@reddit
Not anywhere now. PCH is on chip for AMD and Intel since like a decade or longer?
jenny_905@reddit
Not always. Got a Comet Lake laptop in front of me, has a separate PCH.
You do see it on-package on Intel U series laptop chips though I think.
Straight_Loan8271@reddit
No? The northbridge (memory controller hub) was integrated into the CPU by Intel when Nehalem in 2008, following AMD who'd done it with K8 in 2003. In both cases the southbridge remained outboard and has only since been fully integrated in some mobile designs
Johnny_Oro@reddit
Hasn't been for a long time. Since haswell I think?
Ok-Parfait-9856@reddit
Ahh good to know, guess I didn’t keep up with the times. I thought the chipset was built into the mobo. Guess my question made someone angry too…
Johnny_Oro@reddit
That's unfortunate. Have my up upvotes for asking.
Exist50@reddit
Mobile vs desktop.
EmptyVolition242@reddit
I wouldn't be surprised if they've still got a contract with TSMC that forces them to buy some older nodes. They could've used Intel 4/3.
Johnny_Oro@reddit
Nah, Intel 4/3 are too good for PCH die, and they're occupied with Granite Rapids and the likes anyway. Intel 7 is still busy keeping up with Raptor Lake demands which are still really high due to DDR5 prices. Alternatively they could use older internal nodes, like what is used by Alder Lake N, but it sure is quite outdated and less power efficient. N6 just makes the most sense. Intel's not making anything but i/o dies in that fab, I doubt the utilization is anywhere close to maxed out. And they're not "forced to buy" it, they already own that fab capacity since they bought it.
ComplexEntertainer13@reddit
Not sure it's just raptor tbh. Intel's announcements about 7 shortages. Came around the same time period that we closed in on last date to order SPR.
If they got some late large SPR orders from customers that will be delivered this and next year. That would also explain why Raptor went from fire sale mode to Intel even hiking prices in the spawn of 6-9 months.
Exist50@reddit
It's not that they're too good for PCH. They're moving that way in the future to bring more wafers in-house. Rather, N6 is really good for the task. IO-optimized requires some work that Intel hasn't bothered with in modern nodes. Arguably, Intel 3 would be a step backwards.
But the main reason in the short term is to share IP with LNL/PTL/etc. IP they probably bought externally, in many cases, and this isn't available in Intel nodes.
R-ten-K@reddit
You're thinking of AMD's contract structure with GoFo after the spin off.
TSMC and Intel would most definitively have not structured a contract like that, at all.
Geddagod@reddit
The platform controller tile, eyeballing it, looks to be like a third of the total silicon area. I wouldn't call that small, though you are right it's prob extremely cheap.
6950@reddit
The 18A die is roughly 80mm2 so if it's really a 3rd than it's 27mm2 N6 extremely Cheap
monocasa@reddit
Not a third of the 18A die a third of the total area, or about half of the size of the 18A die.
dropthemagic@reddit
And it will probably be way shittier than the competition
mrblaze1357@reddit
I just ran benchmarks on the X7 368H for my work. Battery life is similar to a Ultra 7 268V, single core is about on pare with a Ultra 7 265H, and Multi core is similar to the 265HX. Not bad. Couldn't get good benchmark scores from the B390 iGPU due to driver issues. But from just a CPU perspective it's really good.
raulgzz@reddit
The b390 igpu is a TSMC chip.
260X@reddit
Let the 18A+++ era begin!
Seanspeed@reddit
We joke, but Intel really does get those evolutions into pretty good shape. First gen of a new node is usually a bit disappointing, but by 2nd and 3rd gen, they're usually cooking. Obviously a new node process is usually better, but in these days when things are taking longer and getting more expensive, I think there's actually quite a bit more room for these evolutions.
DehydratedButTired@reddit
Stop cheering on the bandaids. Intel is finally investing like they should after years of cutting research budget. The +++++ process was only ever stemming the bleeding and setting themselves up for an even harder future. They need to get their smaller node process down and it needs to be reliable or they will still fail.
Seanspeed@reddit
I'm not 'cheering on the bandaids'. What a ridiculous interpretation of my post.
Also, even TSMC are basically doing like 3nm++ these days, they just dont call it that. smh
raulgzz@reddit
It's not the same as intel back then. With tsmc you see real improvement as you can see Apple m3 to m5 are 3nm chips and you can clearly measure massive improvements in performance and efficiency.
III-V@reddit
There's really no reason why they'd be stuck on 18A. They got stuck on 10nm, and consequently, the 7nm that derived strongly from 10nm, because there were a ton of changes at that node. 14A is pretty straightforward from 18A. Who knows from there, though.
Exist50@reddit
Well Intel 4/3 was also supposed to be very straightforward, but they screwed that one up too.
WhoTheHeckKnowsWhy@reddit
and who knows; perhaps they can make the i3 series on plane old intel 3.
Exist50@reddit
Because the TSMC nodes are much better, and at some point the cost to product sales is too big to ignore. TSMC also, you know, actually let's you ship on time.
Seanspeed@reddit
Manufacturing capacity issues. They only have so many fabs/machines for these super modern processes, so have to prioritize.
But yea, it's disappointing cuz I think something like a monolithic Arrow Lake on Intel 3 probably would have been really good.
Geddagod@reddit
If these new Wildcat Lake CPUs are supposed to be replacing aging budget RPL-U and ADL-N series chips... how exactly is it them reducing dependence on TSMC? The previous products in this slot were made on Intel 7.
If anything this would be them increasing their dependence on TSMC, since the PCT tile is being made externally.
Ok-Parfait-9856@reddit
It’s progress compared to ARL, since that was all tmsc. Having just the pch on tmsc 6, an older and cheaper node, is better than having the whole chip made by tmsc
Exist50@reddit
This chip doesn't replace anything ARL. More like RPL-U in practice.
Ok-Parfait-9856@reddit
Yeah I forgot wild cat lake is pretty low on the performance spectrum. And considering mid-high end NVL will be tmsc most likely, that’s not a great look for Intel unfortunately
Johnny_Oro@reddit
Unlike RPL though, WCL has really good single core performance combined with exceptional power efficiency though. GB6 MT score at 8000 seems to be not far behind ryzen 5 phoenix and desktop ryzen 5600, and above i7 ADL-U, that's no slouch at all. And it will be a much higher volume product than NVL. Also NVL-H (and hopefully some NVL-S too) will have lower end SKUs built in 18A.
Ok-Parfait-9856@reddit
That’s good info, thanks man. The single core performance definitely counts for something, that’s one of the most meaningful metrics for typical laptop use (web, docs, etc)
The average laptop buyer doesn’t need a lot of cores, just 4-6 fast cores that don’t suck power like a vacuum. When you mention RPL, do you mean RPL mobile chips have poor single core performance or desktop RPL too? I know it’s hard to compare mobile and desktop chips
Johnny_Oro@reddit
As long as DDR4 is in production and significantly cheaper than DDR5, ADL and RPL won't ever be replaced though. I don't think WCL "replacing" anything. It's a new segment.
ConsistencyWelder@reddit
So they can only produce the lower end parts in their new factory?
beigemore@reddit
Intel Core Series 3 Pro X Series
amdcoc@reddit
Intel will be the only chip mfg on the planet earth surviving when Taiwan Reunites with China.
iDontSeedMyTorrents@reddit
What happened to Samsung in your scenario?
amdcoc@reddit
Assmung is waaaay behind Intel 18A.
iDontSeedMyTorrents@reddit
Even if we say that is 100% fact, customers are still preferring to take their business to Samsung over Intel.
amdcoc@reddit
Customers are always late to Winds of change.
hardware-ModTeam@reddit
Thank you for your comment! Unfortunately, your comment has been removed for the following reason:
Please read the the subreddit rules before continuing to post. If you have any questions, please feel free to message the mods.
Terrh@reddit
What about Gloflo, TI, Micron, mitsubishi, Renasas, Sony, ON Semiconductor, Samsung, Infineon, Broadcom, and like, the 1000 other chip manufacturers that have enothing to do with taiwan or china?
Slasher1738@reddit
Rapidus and Samsung too also, it's not impossible that TSMC US breaks off from TSMC Taiwan. That's why there's a massive expansion plan underway
amdcoc@reddit
Assmung is waaay behind, even MTK won’t fab there 😂👌🏻😂
Slasher1738@reddit
Braggers can't be choosers when the situation is critical
amdcoc@reddit
And that’s why Intel is the most undervalued company in the world at this point of time.
CobaltFermi@reddit
Of late, Intel had been using TSMC to fab many chiplets on their products. With 18A panning out, the goal will be to manufacture the entire product in-house.
Exist50@reddit
They're going back to TSMC for NVL compute though. So mixed bag to be generous. And if Intel Products is actually independent, they might even use less Intel going forward, since corporate pressure is what kept them away to begin with. Big "if", though.
SniperChicken39@reddit
Yeah we've known for 4 months