Intel's 18A Node Outperforms TSMC N2 and Samsung SF2 in 2 nm Performance Class

[kondamin]

. It's probably also why Alchemist and Battlemage are made by TSMC; GPUs benefit more from density. Smartphone SoCs and GPUs and many server CPUs favor density over performance, and this makes it hard for Intel to sell its performance-optimized nodes to most customers.

so ibm power on intel

 

[Vayra86]

Seeing is believing. Theory != practice.

 

[Prima.Vera]

Anyone has the plain English translation to all that technobabble please?

 

[Tek-Check]

Anyone has the plain English translation to all that technobabble please?

Intel has another hope for their 18A node on Panther Lake.
They have had a few other hopes in the past, but it turned out that they continued to lose, gradually, market share across all PC and data center segments in recent years. That's your translation.

 

[Wirko]

With TSMC you could build toward area (N2), or toward increasingly higher performance/area/power (N2P/X) and then even better at higher cost (A16); similar with Samsung (eventually).
With Intel it would appear you would be stuck with a very particular parameter, catered toward high performance but not factoring in optimal cost/area/power for a product.

Hence I think this is a fairly odd way to judge things, as not all designs will fit perfectly within the tuning of 18A, nor will/would they have an optimized cost (for companies nor their clients).
TSMC/Samsung do have those options and/or will.
I would feel much better about article if it relayed the differences between 18A, A16, and N2Z for performance; or 18A, N2 (which is built for density), and 2N (which is in-between N3P and N2) for cost/power/density.

It's fair to give something a rating based on performance (ie clockspeed), as that relative capability is important, but then one should also do it for optimized area/power consumption (which they have not).
Especially when comparing the base nodes from the companies that are not Intel, all of which generally first focus on density/low power rather than performance. Hence this is slightly to highly misleading.

IOW, cool; they can run a BSPD chip at 1.2v+, where-as N2B is denser but isn't aimed toward that; probably ~1-1.05v. Probably similar for Samsung. That doesn't make 18A ~15% better, as this implies.
Also, almost nobody (outside of Apple) actually is using N2B. They will use N2P (~1.15v) or n2x (1.2v+) and/or A16 when available and feasible (especially in cost) versus N3(E/)P (and perhaps Samsung base 2nm).

If companies are considering using 18A, they would likely also consider A16 for the advantages of BSPD...but even then TSMC may cater designs toward different mixtures of density/performance.
Considering the cost, many may not, and instead opt for something like N2X (as AMD appears to be doing; high perf [perhaps similar to Intel] but similar/slightly less density and/or similar/worse power consumption).

Include A16 and SF2 in the comparison? Maybe but what are the roadmaps, are the timelines comparable?

However, while everyone in this thread talks about power/performance/area and perhaps yields, you're the only one who has mentioned cost. This matters more than ever before as the price per transistor is probably going up on 2nm-class nodes, not down, compared to previous nodes. These nodes will be of interest to those who need, and can pay for, density. Server CPUs and AI accelerators rather than desktop and mobile chips.

 

[Zodwraith]

I remember the hype surrounding Intel 20A and that didn't turn out well.
Personally I'll only believe such claims when mass manufactured products are benchmarked by independent third parties.

There was nothing wrong with 20A. It was cancelled only because 18A was so close behind it made no sense to continue pouring R&D money into both processes. If they'd have continued with 20A it would just be taping out now, mere months ahead of 18A.

Keep in mind 18A isn't just a slight density increase, it's new technology with back side power delivery and a new gate design. The only way keeping 20A alive would have made sense was as a 3rd party node, which obviously no one showed interest in. It was smart to cancel 20A and go all in with 18A.

 

[Squared]

There was nothing wrong with 20A. It was cancelled only because 18A was so close behind it made no sense to continue pouring R&D money into both processes. If they'd have continued with 20A it would just be taping out now, mere months ahead of 18A.

Keep in mind 18A isn't just a slight density increase, it's new technology with back side power delivery and a new gate design. The only way keeping 20A alive would have made sense was as a 3rd party node, which obviously no one showed interest in. It was smart to cancel 20A and go all in with 18A.

But how do we know that 20A wasn't just behind schedule and was renamed 18A? Someone suggested that possibility once and I said that if Clearwater Forest comes to 18A on schedule then it's the real 18A because Clearwater Forest needs high density libraries that 20A doesn't offer. But now Clearwater Forest is delayed, and there's an "18A-P" in the roadmap, and Panther Lake is using 18A and Panther Lake doesn't necessarily need the high density libraries. (Meteor Lake doesn't have access to them on Intel 4, but Sierra Forest does on Intel 3.)

 

[tfp]

But how do we know that 20A wasn't just behind schedule and was renamed 18A? Someone suggested that possibility once and I said that if Clearwater Forest comes to 18A on schedule then it's the real 18A because Clearwater Forest needs high density libraries that 20A doesn't offer. But now Clearwater Forest is delayed, and there's an "18A-P" in the roadmap, and Panther Lake is using 18A and Panther Lake doesn't necessarily need the high density libraries. (Meteor Lake doesn't have access to them on Intel 4, but Sierra Forest does on Intel 3.)

Both 20A and 18A are behind schedule. 20A was not renamed 18A, because 20A was not going to have back side power deliver and the new gate design. If Intel was lucky they save some investment because they are not needing to improve yields on "similar" nodes for 2 different gate designs.

 

[ncrs]

There was nothing wrong with 20A. It was cancelled only because 18A was so close behind it made no sense to continue pouring R&D money into both processes. If they'd have continued with 20A it would just be taping out now, mere months ahead of 18A.

Keep in mind 18A isn't just a slight density increase, it's new technology with back side power delivery and a new gate design. The only way keeping 20A alive would have made sense was as a 3rd party node, which obviously no one showed interest in. It was smart to cancel 20A and go all in with 18A.

Intel has shown Arrow Lake made on 20A. If there was nothing wrong with the 20A process they would've shipped it instead of wasting hundreds of millions on porting Arrow Lake to TSMC which is really supply constrained. So much so that there's still no mainstream Dell OptiPlex or Inspiron based on Arrow Lake - Intel simply can't ship enough for the massive requirements of large OEMs like Dell or HP.
There were rumors that lower tiers of Arrow Lake would be made on 20A while high end was TSMC which would alleviate the problem I described above. This however did not happen, so I do not believe 20A was viable and instead was a huge loss (R&D for the node itself and potential sales of high volume products) for Intel.

 

[Squared]

Intel originally promised 18A would be "manufacturing ready" by the end of last year, but it entered "risk production" just recently this year, so it's a little behind. (However, it does appear to be ahead of TSMC's N2 schedule by a good bit.) But when 20A was canceled, Intel claimed it was because 18A showed "early success". How can 18A be 4 months behind schedule if it's showing "early success?"

 

[tfp]

Because 20A yields where not good enough to continue and behind schedule as well, though they did show that backside power did work on 20A (and gate all around?)

The early success statement related to where they were in the over all process is just disconnected with the actual calendar. Both statements are "true" but don't hold the same meaning as a calendar statement would.

 

[SirJynx]

I remember the hype surrounding Intel 20A and that didn't turn out well.
Personally I'll only believe such claims when mass manufactured products are benchmarked by independent third parties.

I'm not sure what you mean by that, 20a was LONG cancelled before there was such a thing as hype for it.

Will believe when I see it, lol.

Looking forward to Intel releasing anything on their 18A in 1H26.

Intel will have 18a products 2h25, the mass production and foundry availability is what will ramp in 1h26

So do I, but I also think not figuring in TSMC/Samsung's BSPD (A16/2z) processes is slightly disingenuous; where BSPD improves area 10-20% and/or performance by 5-10%, which would even things out.
And in reality may yield (especially for outside customers) in a similar time-frame. Just using the '2nm' nodes because of the name is a little odd. Why not 20A for Intel, then? A16/2z is just 2nm w/BSPD.

How much do you think their rating would change from the addition of a BSPD network for TSMC/Samsung? Now add, for instance, TSMC's density improvement (options) vs Intel. It should equal out.
Samsung perhaps slightly behind, but they're also generally a value contender, so it makes sense. They'll also likely be close-enough for similar designs to make sense, and may gain most from BSPD.

With TSMC you could build toward area (N2), or toward increasingly higher performance/area/power (N2P/X) and then even better at higher cost (A16); similar with Samsung (eventually).
With Intel it would appear you would be stuck with a very particular parameter, catered toward high performance but not factoring in optimal cost/area/power for a product.

Hence I think this is a fairly odd way to judge things, as not all designs will fit perfectly within the tuning of 18A, nor will/would they have an optimized cost (for companies nor their clients).
TSMC/Samsung do have those options and/or will.
I would feel much better about article if it relayed the differences between 18A, A16, and N2Z for performance; or 18A, N2 (which is built for density), and 2N (which is in-between N3P and N2) for cost/power/density.

It's fair to give something a rating based on performance (ie clockspeed), as that relative capability is important, but then one should also do it for optimized area/power consumption (which they have not).
Especially when comparing the base nodes from the companies that are not Intel, all of which generally first focus on density/low power rather than performance. Hence this is slightly to highly misleading.

IOW, cool; they can run a BSPD chip at 1.2v+, where-as N2B is denser but isn't aimed toward that; probably ~1-1.05v. Probably similar for Samsung. That doesn't make 18A ~15% better, as this implies.
Also, almost nobody (outside of Apple) actually is using N2B. They will use N2P (~1.15v) or n2x (1.2v+) and/or A16 when available and feasible (especially in cost) versus N3(E/)P (and perhaps Samsung base 2nm).

If companies are considering using 18A, they would likely also consider A16 for the advantages of BSPD...but even then TSMC may cater designs toward different mixtures of density/performance.
Considering the cost, many may not, and instead opt for something like N2X (as AMD appears to be doing; high perf [perhaps similar to Intel] but similar/slightly less density and/or similar/worse power consumption).

This article is very nuanced, and kind of comparing apples to oranges imho, and comes across trying to put Intel in the best light possible using a very particular metric and comparison(s).


Bingo. This guy gets it.

I don't see how it's disingenuous, the timeframes are not the same. Intel already has revisions of 18a, 16a, 14a on the roadmap that those products can be compared to. They can't use 20a from Intel as that was already a scrapped product to focus budget on 18a ramp instead. Intel isn't yet using their High-NA EUV lithography machines until 16a, of which I believe they bought out the first 8 from ASML and will have at least a short term exclusivity with the technology, and that's were we might see Intel really pull ahead with it's process leadership goals more commandingly.

Here we go again with Intel lithography propaganda. They haven't been at the top of their process gaming in like a decade. If they release desktop and laptop CPUs that use 18A soon (before TSMC beats them) say by late 2026 it'll be great but who knows what's coming. Their competition isn't resting either.

18a doesn't use Intel's new lithography machines. Those are for 16a. Backside Power is done by a different machine that grabs the wafer, flips it over, and does a cycle of polishing and washing to reveal the power... on the backside

TSMC has backside power delivery slated for A16, which according to TSMC and Intel roadmaps will land around the same time as Intel 14A*. TSMC might ramp up faster and produce a greater volume, but if the first Intel 18A products are released this year (as promised) and feature backside power delivery (which Intel has not promised) and the rest of the roadmap is true, Intel will have had backside power delivery for about a year before TSMC.
View attachment 395081View attachment 395083

*Intel's roadmap also puts 18A in 2024. It just entered risk production and I believe the current year is 2025. TSMC's roadmap has N3E in 2024, and Apple M4 chips on N3E did reach consumers that year. So TSMC's roadmap uses consumer-reaching dates and Intel's uses another metric which is perhaps 1 year before consumer release.

That's an older roadmap, as 20a was scraped. But 18a is slated for 2h25 and it will have BSPD. We'll see it in use by foundry customers by 1h26. Intel will always be customer zero for new products

If so that's good news.

I guess the question I have then is why is Intel subbing so much fab work to TSMC?

It let them learn the customer experience with TSMC to prep for foundry, it allowed them to invest more heavily in all the new technologies they're finally bringing to market over the next few years, it let them use TSMC as a stop gap to not split investment into current manufacturing they had already fallen behind on to TSMC, which is allowing that equipment to be converted for use with foundry customers who don't need cutting edge chips like 65nm Intel 12 with UMC, and will now also allow the products team to make the best possible products possible by letting them mix and match products with TSMC, so Intel doesn't have to compete on every segment with them. TSMC has significantly more capacity after all. You can't just compete with a titan like that overnight.

Also, "5 nodes in 4 years" has been the most meaningless and stupid narrative ever. The reality has shown that such propaganda pitched to tech media is not reflected in what's happening on the ground. Intel has gradually and literally lost market share with all those nodes in every single segment of PC and data center industries. The loss in desktop has accelerated in recent quarters more than before, and in data center AMD and ARM designs are going to reach around 50% of share within one year, if not sooner.

In addition, Panther Lake might have one or two showcase laptops by December 31, just like they did with Meteor Lake, but almost nobody in the world will be able to buy it in any meaningful quantities before mid Q1 2026.

The point of 4 nodes in 4 years was to make a roadmap on how to get to process leadership on an incrediblely short timeframe. Manufacturing chips is not something you can just "fix" in a few months. It takes years of research and execution and tooling and building out and training, etc.

If those nodes remained somewhat competitive, why chip designer companies do not produce their leading chips with Intel Foundry, a branch of business that has experienced heavy, multi-billion losses, quarter after quarter after quarter? Surely, Apple, Nvidia, Qualcomm, MediaTek, AMD and others could ask Intel to etch their leading chips, no?

Do they trust Intel Foundry to deliver on what they need? The simple answer is no. And that's the key problem with Intel Foundry business. Intel has had a fundamental conflict of interest between branches of their business. Intel Foundry is yet to produce a leading chip for non-Intel entity that is better than Intel own chips.

TSMC doesn't have this issue as they are simply the leading global foundry without such conflict of interest.

Yes, but they haven't outsourced entire generations of front line products, such as Lunar Lake. They were forced to do this as their own nodes were crappy at that time. TSMC has a better product and even the US government asked them to help save Intel Foundry from further bleeding.

You have a misunderstanding of Intel Foundry. Intel had the capacity to build chips for Intel at the time they decided to spread out to foundry. The reason Intel is losing so much money, is because they've essentially done nothing but build out just to have the capacity to make chips for other companies, create new softwares and sales interfaces, create parameters and earn foundry certifications, build test wafers and samples for customers, etc. you can't just flip a switch and sell products to other people when you have your own to sell. This was always going to be the timeline. Mostly all Intel chips lately are made by TSMC, so Intel can focus on retooling older tools for new customers, phasing out their end of line products. Building new fabs to house their new technologies for both them AND new customers for the first time, pioneer new assembly technologies (which is the only real foundry product leaving their doors atm), and ramp up for this new Intel basically. The earliest foundry customers they got were the department of defense and AWS, and neither of them has had a chip yet, because all the big customers are signing up for 18a, which means the opportunity of profits doesn't really exist yet. Pat always said 2027 is probably the timeline for foundry to make lots of money, because e everything else is invest invest invest. Lip-bu might rebalance that a bit.