Intel announces 18A process node has entered risk production — crucial milestone comes as company ramps to Panther Lake chips
Posted by III-V@reddit | hardware | View on Reddit | 54 comments
-protonsandneutrons-@reddit
How on Earth can Intel still claim 5 nodes in 4 years is "alive"? I didn't hear 5N4Y in the presentation; if it's just Tom's Hardware, that's somehow worse—why run defense for a mega-corp?
No, Tom's Hardware. Because Intel cancelled 20A (and lost a major, widely-publicized Foundry customer on 20A), "5N4Y" is unequivocally dead. And it was never much alive as Intel counted half-nodes as full nodes.
We can do without the spin.
pianobench007@reddit
Pat talked about how Intel 10nm fell behind and we kept having 14nm in 2016, 2017, 2018, 2019 and 2020/2021. They were supposed to have a 10nm ready by 2018/2019.
So they were delayed about 2 years. 10nm finally launched around 2021/2022 then we saw Intel 7 for consumer products and Intel 4/3 for data center.
Intel 20A was skipped due to better yields on Intel 18A so they decided to accelerate 18A process. 20A and 18A are just names. Labels. Marketing. Both technology processes featured RibbonFET transistors and PowerVia. Two technologies for the process nodes.
As an example. TSMC 16nm was their first finFET transistor process node. Intel introduced finFET transistors first for 22nm. And TSMC 20nm featured planar transistor for their node technology. So 20A or 18A is just a marketing label for a process node technology. Not much changes between 20A and 18A besides new and improved techniques.
Pat did mention time and time again that the reason Intel failed so badly at 10nm was specifically because they had no backup for their process technology. They became too cocky and sure of themselves as to NOT have any backup manufacturing process in the pipeline.
20A was expected to be mfr. with EUV. 18A is expected to be manufactured on high NA EUV. So this time around Intel has its backup plans. Same for Intel 4 and Intel 3.
From what I've seen Intel 3 is doing just fine. Better yields for both datacenter chips.
https://newsroom.intel.com/opinion/continued-momentum-for-intel-18a#:~:text=when%20we%20set%20out%20to,may%20be%20required%20by%20law.
Exist50@reddit
Intel 4 was late about 2 years (originally planned for '21, shipped very late '23). 20A/18A are also about 2 years late.
That's just false. 20A was skipped because it was too broken to use for a product. Same thing happened with 10nm. Notice how 18A was also downgraded to 20A performance.
18A hasn't been entertained to be on high NA in years. Also, two failed nodes isn't exactly the best example to use for a turnaround in strategy...
pianobench007@reddit
I linked intel statement on why they moved forward with 18A. It is from their newsroom.
Exist50@reddit
At no point in that link do they claim defect density on 18A to be better than 20A as your comment did. And you should know by now not to take Intel's statements at face value.
pianobench007@reddit
I don't have data on any of that as I am not a leaked or have inside knowledge. But what I do know is that they've bought the expensive EUV machines and the high NA EUV machines along with they have experience making chips. Foundry we don't know yet as they don't have a track record of course.
But I figure if TSMC can do it while Intel was in the lead for finFET, Intel can catch up again with ribbonFET too. It's not like there are patents on who can make those fets. At the end of the day it just takes constant drive for success ! And I am sure that both sides of the pacific pond have it.
Its just a node name anyhow. And my word is nothing versus yields that their customers want. They know how much product they can get with each team and will do an internal calculation and pick one that makes the most sense. That's how I see it anyhow.
Intel can also just sell their node at a loss which is another option to.
Exist50@reddit
Machines have never been Intel's problem, even though they've tried to use them as a post-hoc rationalization. Intel 10nm/7 and TSMC N7/N7P were made with the same DUV machines, but TSMC had no problem making a commercially successful node, on schedule, while Intel could not. There's more to running a foundry than the lithography equipment.
I worry that the circumstances that caused that reversal haven't meaningfully changed since. At a most basic level, Intel Foundry remains broken so long as they cannot hold to a schedule. Certainly the pattern even for Intel 4/3 and 20A/18A has not been encouraging.
They do, but given that Intel has almost no 3rd part foundry customers, then well, it paints a grim picture. Though the problems aren't just technology. Pat was very bad at attracting customers.
Strazdas1@reddit
Intel 10, 4, 3, 20A, 18A. I count 5.
Geddagod@reddit
20A got canned. Intel 7, 3, and 18A are all "sub node" jumps, or just more complete versions of the main jump nodes, those being Intel 4 and Intel 20A.
Strazdas1@reddit
if 20A is canned, then 18A is not a subnode.
Tiny-Sugar-8317@reddit
Yeah, factually speaking "five nodes in four years" has turned into "four nodes in 5 years". But whose counting? Clearly not TomsHardware.
imaginary_num6er@reddit
More like 4 refreshes in 5 years
Geddagod@reddit
GLC the new Skylake lol
III-V@reddit (OP)
Sorry, Tom's was the only source I found. I didn't exactly look too hard, but I didn't even see it on Intel's website.
-protonsandneutrons-@reddit
Ah, no, no, I'm not blaming you. Just thought Tom's seems to be adding words to Intel's presentation.
I wouldn't have heard this part of the presentation until you shared the link, so cheers.
SlamedCards@reddit
Literally impossible it's in risk for panther lake because it wouldn't launch until something like mid next year and same presentation they said it's launching this year.
Kevin was up there talking about external clients. So gotta wonder if there's 18A for Intel and 18A for foundry clients
jaaval@reddit
Risk production means they are running real product production on it, no longer just test chips. i.e. they think it should work as designed. The risk part in risk production means they have to accept a bit worse results while scaling up production and tuning the pipelines.
Exist50@reddit
No, not for production silicon. Intel's launch schedule is about a quarter off production. So that won't start till Q4'25.
ThankGodImBipolar@reddit
So that “H2 2025” launch for PTL will probably look similar to the MTL launch?
ProfessionalPrincipa@reddit
I think that's the safe bet. The MTL, SRF, GNR, LNL, and ARL-U launches have all gone like that. It ain't like the good ol' days when they had boatloads of stuff ready to go on the snap.
Geddagod@reddit
Am I tripping? I swear LNL's launch was actually decent, and we saw decent volume in 2024.
I agree MTL was bad, SRF and GNR appeared to be paper launches as well, and is ARL-U even out yet? Lol. But LNL?
ProfessionalPrincipa@reddit
What's decent? Intel claims they shipped over 1.5 million LNL in 2024. Which in the grand scheme of things isn't a huge number.
Geddagod@reddit
Ehhh I'll admit I was just going off a vibes thing based on what I saw on bestbuy in stock lol. I was looking into buying a LNL laptop last year. Basically being able to get a couple good different LNL selections in stock.
eding42@reddit
Also Arrow Lake-U is just rewarmed Meteor Lake, it'll probably trickle out soon as soon as Meteor Lake stock dries up but the actually interesting Arrow Lake mobile launches were the H and HX chips, and those are out already.
Geddagod@reddit
I think ARL-U is interesting, maybe not from a consumer POV, but to see how much better the Intel 3 tile is vs Intel 4 tile.
eding42@reddit
I agree. Should be slightly better.
eding42@reddit
To be fair Lunar Lake was a late-Summer launch, yes there was some stock right away but most SKUs launched in December/January. I'm guessing we might get one halo Panther Lake SKU like December 31st with reviewers, with the bulk of the volume by Feb
Exist50@reddit
If Intel's official calling it a '26 product, safe to say even token availability in '25 isn't happening.
ThankGodImBipolar@reddit
Lip Bu Tan’s letter from the last shareholder meeting says “in the second half of this year… launch Panther Lake,” which is what I was referencing. Silly me forgot that they don’t need any actual product to “launch” their lineup…
Exist50@reddit
Yup. "Launching now with availability Q1'26".
Helpdesk_Guy@reddit
Yes, the usual paper-launch with no availability on 20–31. December, to meet legal time-lines …
SlamedCards@reddit
Risk production is normally ES type stuff. Intel has been doing ES steppings since last year for PTL
xternocleidomastoide@reddit
Not necessarily. Risk production can be past ES, and includes commercial samples in field.
HorrorCranberry1165@reddit
risk production means they still are making test chips, but at higher scale. Test chips are best way to measure health condition and quality of process.
my_wing@reddit
This is not what it meant, risk production means that they are producing production chip. The Yield is the one they needed to solve.
We are now in April, it takes 5 - 8 weeks (from silicone disk to fully packaged chip).
If that is 8 weeks instead of 5 (totally possible as 18A is complex), then it will be June that it will be at the packaging plant. Then it will needed to install into Acer, Asus, Dell, HP, .... and it will also take 4 weeks for production. That will already be July Aug, it will take 2 weeks to be in Best Buy, JB Hi Fi (Australia) ... So if they don't start Risk Production now, then we can see PTL in Laptop (Assume PTL is Laptop first). So it is old news, because they already said PTL is going to paper launch Q4 2025.
xternocleidomastoide@reddit
Different foundries have slightly different definitions for what Risk Production means.
This basically means that Intel has at least done a full bring up of a design on 18A. This is the entire flow for the process is up and running pretty much.
Exist50@reddit
Because Intel has consistently misled the press as to the state of their process nodes. It's really that simple.
ElementII5@reddit
We have been farming negative karma here for years telling them how it is. Yet still surprised faces everywhere.
Geddagod@reddit
Please don't try to group yourself with him
"Zen 5 Turin is gonna have 2x the perf/watt of GNR and Zen 4 sever CPUs" lmaooo
ElementII5@reddit
I think it came out as 1.4x. not too bad for napkin math. What I didn't take into account was how much Intel jacked up power consumption to be competitive.
Geddagod@reddit
It came out to 1.2x from the phoronix review. The 2P GNR system was plagued by a scaling bug that Intel hasn't fixed yet, so comparing 1P vs 1P would be more accurate.
And it came out to 1.2x using Phoronix's test suite, which has its own set of problems. Using specint2017, which mind you isn't perfect, but is the industry standard and better than Phoronix's aggregate average, you would see that AMD themselves admit that Intel's GNR scores esentially the same with the same power draw as Turin (ISSCC 2025).
2x vs 1.4x is pretty bad, even for napkin math. And in reality, 2x-1.5x vs 1.2x (or even just <1.1) is even worse.
The worst part is that I don't even think you ever explained your napkin math either. All you said was see in you 2 weeks, where you turned out to be proven wrong anyway (though that didn't stop you from trying to spin it the other way).
This is just outright wrong in 2 ways. First of all, Intel didn't jack up power consumption to be competitive at all.
From Phoronix's review, the MRDIMM version of GNR uses like a whopping 8 more watts than Turin on average, for a <3% difference, and the regular memory GNR literally has a lower reported average power draw (while Turin still only scores <1.2x better mind you) by a miniscule amount.
Secondly, even if Intel jacked up power draw to be competitive in perf, that would widen the perf/watt gap, since increasing clocks and power leads to worse perf/watt due to how power scales.
You are right, I should quote you correctly.
Here.
Actually, it gets better, because the chart in that post uses you citing a 2.01x gain actually, not just a 2x gain. My bad.
SlamedCards@reddit
Could you share paper name of thurin vs gnr
Geddagod@reddit
It was just AMD's presentation about Zen 5.
Here (1:01:13)
SlamedCards@reddit
Interesting. That's quite a good indication for Intel 3 no?
Maybe 5-7 ish precent gap between the 2 for spec int (does have faster ram). And redwood cove has not changed from golden cove much
Tbh I thought gap was more like 15-20%
So if lioncove was on Intel 3 it'd be pretty good performance (tho high power consumption since Intel 3 isn't as dense)
And DMR skips to panther cove over lioncove and it's derivatives.
I also saw you mentioned A16 is something like 7-10% bump. Ik you don't buy into theory of 18A having similar PPA to N2. But if you did, you would say Intel essentially claims 10% ish uplift for 18A-P then it would be performance level of A16 in very late 26. Compared to maybe late 27 products for A16
Geddagod@reddit
An ok one, a decent N4 competitor it looks like, ignoring RWC's much larger area (a decent percent of which is prob contributable to Intel's worse design/layout team).
It prob will depend on the benchmark used too. I can totally see Turin outperforming GNR by that margin in many workloads.
So LNC almost certainly faces a very similar problem that Zen 5 did vs Zen 4, tock cores usually don't improve in perf/watt much at the power ranges per core that server CPUs use.
I lowkey butchered that sentence, but esentially, Zen 5 at 3-2ish watts per core doesn't score much higher than Zen 4 at the same power draw in specint 2017.
I would wager LNC would face a similar issue. And unless Intel would also increase the TDP of a LNC-GNR to take advantage of the extra perf headroom a tock core presents, we wouldn't really be seeing that advantage.
Yes, but DMR would also have to face off against Zen 6 server skus.
The tock core in DMR would likely have to face off against a core that has almost as high IPC in Zen 6 Venice. Maybe DMR's P-core has a 5%ish percent lead.
However I do want to point out that SPR had a core that had like a 15-20% IPC lead vs Milan, and it still did just as good or worse as Milan did. For sever skus especially, the uncore and perf/watt at lower wattages per core matters a lot.
For HPC skus only*
That bump is over N2P, not standard N2. So N2P is a 5-10% perf/watt bump over N2, and A16 is a 8-10% bump over that.
TSMC claims volume production in late 26, however seeing what is rumored to have happened with N2, late 27 could be when A16 products actually do come out.
ElementII5@reddit
These are Server parts. You can't just compare SKUs. Perf/Watt is the important metric. See this comment of mine.
https://www.reddit.com/r/hardware/comments/1g0pa1d/phoronix_amd_epyc_9755_9575f_9965_benchmarks_show/lrarfru/
Geddagod@reddit
So I'm unsure what you mean by this. You can't just compare SKUs? I'm assuming that this is you complaining about me not using the 2P vs 2P results...
First of all, the 2P vs 2P results are bugged for Intel. In my reply to you a couple months ago, I think I was speculating, but since then, the author of the article himself confirmed it.
Second of all, comparing 1P vs 1P even for server is still fine. AMD themselves do it.
That's still what I was comparing?
Well first of all, I actually linked that comment you just linked twice in my previous comment.
Second of all, I saw that comment, and replied to that comment, months ago too. It's just complete BS the way you compared them.
And ik you saw my reply to your comment too, because you replied with this further down the thread, which I then responded too again.
ElementII5@reddit
How can anybody still be surprised? Well I have been saying so here and been downvoted for it.....
Exist50@reddit
It's funny that in the thread just the other day, you had people swearing up and down that 2026 was fake news. Oh, and this puts 18A realistically behind N2 in schedule, to say nothing of the deficit vs N3.
Sapiogram@reddit
"2026 fake news" as in it's going to be after 2026? Or before?
Exist50@reddit
People claimed it would be in 2025, because Intel kept on saying so. Until they didn't.
ReskatorBC@reddit
Even if they produce a Jupiter Lake version I’m not buying their 💩 anymore. No No No !! Thank you.
my_wing@reddit
If intel is producing the next ryzen, please don't buy it, if intel is producing nVidia 6090, please don't buy it.
AutoModerator@reddit
Hello III-V! Please double check that this submission is original reporting and is not an unverified rumor or repost that does not rise to the standards of /r/hardware. If this link is reporting on the work of another site/source or is an unverified rumor, please delete this submission. If this warning is in error, please report this comment and we will remove it.
I am a bot, and this action was performed automatically. Please contact the moderators of this subreddit if you have any questions or concerns.