News Intel’s turnaround plan hinges on this one chip family – Clearwater Forest pictured, Intel’s first 18A chip slated for high-volume manufacturing.

[thestryker]

18A is certainly going to dictate the future of Intel and CWF being on time and meeting expectations should speak volumes to the industry. I find it very interesting that they're going big right out of the gate as that would seemingly indicate confidence in the node.

 

[Guardians Bane]

In the good old days, MetalMessiah would already be making several corrections to this article.

Am i the only one who kinda misses that guy?

Omg ... That guy was always so damn negative. I loved seeing how hard he tried to bash authors and failed! He was a laugh!

 

[bit_user]

But I never ran OEM or recommended settings from anyone.... They were all smokin crack from the top down! Nobody will ever convince me that a chip running so hot that it can literally boil water, is a safe operating temperature. So I never let mine get that hot. I tuned my system by temperate, not by any other metric. Dumping more power into these chips do not make them run any faster, it just makes them hotter! From day one I lowered the thermal limit on this chip and set my power limit @ 220w if memory serves. I have always run well below the 253w recommended by Intel.

There's been lots and lots written about this, so I won't delve into the details here. However, it's been shown that you don't even need to run at high temps or the 253 W PL2 for Raptor Lakes to degrade to the point of encountering errors. The Xeon E-2488 has a PL1 of just 95 W and PL2 of only 181 W, yet people could quite reliably get them to fail just by running medium-duty server workloads on them, even with down-clocked memory!

 

[usertests]

Hate to break it to the haters and there are many in recent years with AMD's resurgence, but Intel isn't going anywhere. Why you may ask. Just like how General Motors was deemed too big to fail, Intel is as well. The US won't be left without the ability to make tanks and electronics. They're national security infrastructure. For Intel to fail completely, we may as well cut our own throats. We don't lose our lives if AMD and Nvidia fail, some people lose money in the market.

Intel doesn't need to exist in its current form, which is why there's lots of talk about spinning off the fabs, Qualcomm acquiring pieces of it, etc.

Taiwan getting invaded could in theory be a boon to the U.S. fabs and bring in customers who are currently skittish of Intel, but the whole supply chain isn't here so it would just become a miserable experience for everybody. Make sure to buy up $100 computers when you see them on clearance.

Omg ... That guy was always so damn negative. I loved seeing how hard he tried to bash authors and failed! He was a laugh!

He succeeded, often, because of how Tom's Hardware is. Maybe you didn't notice.

 

[bit_user]

There is more than Clearwater Forest to Intel's fortunes.
Sure a 288 Raptor cove equivalent core chip sounds great,

No, it's E-cores. They're Skymont and they're only comparable on IPC to Raptor Cove when measured in single-thread scenarios. If you measure real-world single-thread performance, Raptor Cove is still faster, because it clocks much higher. And as soon as you put a second thread on Raptor Cove or try to scale up, you'll see a CPU with an equivalent number of Skymont cores falling behind, because the Skymont cores are still arranged in quad-core clusters with shared L2 cache and interconnect ports.

That's what's so problematic about your sly use of their Raptor Cove IPC equivalence. They're not the same and it's really disingenuous of you to talk as if they are.

But speaking of those beefed up e-cores, is there any news on how much power they draw and performance they have when the p-cores are turned off yet?

It wouldn't be very relevant, because Lunar Lake is made on TSMC N3.

 

[Dingledooda]

Didn't Broadcom recently complain about how disappointing they found first test drive productions of their chips on Intel's 18A process?

It wasn't tests of their chips it was wafers with test patterns of structures they typically use in their designs and no they didn't. Reuters did an article saying they were because apparently "three sources familiar with the matter" told them.

Were any of those sources from Broadcom to actually know what was going on in Broadcom? I would think if any were then Reuters would of said as such since that would add credibility to the article but no it was just people who were "familiar with the matter" what ever that means. For all we know it's just "Tech Analysts" speculating about what they think Broadcom thinks because of other stuff from Intel and Broadcom that they know.

All Broadcom has actually said is they are "evaluating the product and service offerings of Intel Foundry and have not concluded that evaluation".

 

[rluker5]

No, it's E-cores. They're Skymont and they're only comparable on IPC to Raptor Cove when measured in single-thread scenarios. If you measure real-world single-thread performance, Raptor Cove is still faster, because it clocks much higher. And as soon as you put a second thread on Raptor Cove or try to scale up, you'll see a CPU with an equivalent number of Skymont cores falling behind, because the Skymont cores are still arranged in quad-core clusters with shared L2 cache and interconnect ports.

That's what's so problematic about your sly use of their Raptor Cove IPC equivalence. They're not the same and it's really disingenuous of you to talk as if they are.


It wouldn't be very relevant, because Lunar Lake is made on TSMC N3.

It will be Darkmont, not Skymont but that doesn't mean much since we don't know the difference yet. And we don't know how fast 288 cores in a package will be able to clock but I'm pretty sure Darkmont will be hitting a higher clock than Raptor Cove given the same number of available watts per core. Take 2w, for example because 576w sounds like a plausible guess for a 288 core TDP. As far as the interconnects, Epyc made it work.

I don't think my comparison is misleading because the known comparison is IPC and not IPS, and I did not claim that a server would be running 288 cores at 6GHz. Nobody expects that. I was making the point that these new e cores aren't weak like xeon phi cores or Raptor Lake e cores. They are up there with the current big cores and if you see 3GHz out of them there will be the performance of 288 big cores at 3GHz coming out of that chip package.

 

[hotaru251]

betting the entire company on soemthing its foundry side can do is bad business imho....

 

[sjkpublic]

It is so sad that Tom's Hardware authors are now posting marketing posters instead of details. I do not see much in this article about the specifics. I see just hype. I can post hype too. Just pay me money as what this silly author gets.

 

[bit_user]

I don't think my comparison is misleading because the known comparison is IPC and not IPS, and I did not claim that a server would be running 288 cores at 6GHz.

Clockspeed was only one issue. I listed several others. One was L2 cache contention. Another was interconnect bottlenecks. Yet another was hyper-threading, which yields a further 30% performance advantage per core, or 20% at same power.

​

Again, kinda sus that you only picked one point on which to defend your positioning. I only mentioned absolute single-thread performance for completeness, not because I thought it would be relevant in this context.

I was making the point that these new e cores aren't weak like xeon phi cores or Raptor Lake e cores. They are up there with the current big cores

"Current" as in Raptor Cove? Those aren't really current, any more. More to the point, server CPUs don't even use those P-cores. They use a modified version with an additional AVX-512 FMA and AMX. So, they're still not equivalent even to Sapphire Rapids or Emerald Rapids P-cores.

 

[Amdlova]

From the 13600T to 14600T See some diferences on allocation of Power.
The 13600T you can set the desired power and it will obey 5w to 110w.
The 14600T in other hand you need to mess with all configs to set the power to level beyond of specs. And the cpu perform worse in some bench marks.
Intel made some corrections how the cpu behavior when you set the power...
some Benchmarks the cpu can't pass the 84w wall but has 92w on pl2.

I don't have any cpu with more TDP to test. Like to keep noise down (noise floor) and listen the motherboard making coil noises.

 

[rluker5]

Clockspeed was only one issue. I listed several others. One was L2 cache contention. Another was interconnect bottlenecks. Yet another was hyper-threading, which yields a further 30% performance advantage per core, or 20% at same power.

​

​

Again, kinda sus that you only picked one point on which to defend your positioning. I only mentioned absolute single-thread performance for completeness, not because I thought it would be relevant in this context.


"Current" as in Raptor Cove? Those aren't really current, any more. More to the point, server CPUs don't even use those P-cores. They use a modified version with an additional AVX-512 FMA and AMX. So, they're still not equivalent even to Sapphire Rapids or Emerald Rapids P-cores.

That's ok
Intel will still have those things in the other line.
The e cores version is probably meant to do the things Epic does.

 

[philipemaciel]

I would feel sad for Intel blues had they not tried to sue and bribe AMD out of existence

 

[pointa2b]

I like AMD's processors more (no need to explain why here), but its important for Intel to pull through. Intel collapsing will most likely cause improvements on AMD's end to slow down and stagnate, just like Intel did before AMD gained more traction. If you want the best hardware with the most aggressive improvements each year, you want healthy competition, regardless of who you prefer to buy from.

 

[Loadedaxe]

These Intel Doom comments make me laugh. I got some good giggles tonight reading them. Intel isnt going anywhere and they aren't going to "implode" and no one is going to take them over. All successful companies have some bumps in the road. FFS look at AMD, they were on the verge of "imploding" and they made it, Intel has more money in their sock than AMD has as a whole.

 

[Mama Changa]

18A is certainly going to dictate the future of Intel and CWF being on time and meeting expectations should speak volumes to the industry. I find it very interesting that they're going big right out of the gate as that would seemingly indicate confidence in the node.

Literally, Panther Lake depends on 18A working too. Panther Lake is Lunar Lake much more performant replacement offering more cores in high-end models. I hope 18A is a huge success. TSMC won't have an answer until N2 or A16. Intel is ahead on GaaFET and BSPD. If Intel pulls off 18A, AMD will suddenly be behind on the process tech. They won't have n2 based Epyc until Zen 6 ships in 2026.

 

[bit_user]

Intel collapsing will most likely cause improvements on AMD's end to slow down and stagnate, just like Intel did before AMD gained more traction. If you want the best hardware with the most aggressive improvements each year, you want healthy competition, regardless of who you prefer to buy from.

There are other competitors, now. Qualcomm, Apple, and soon MediaTek are all building competitive ARM-based processors. There's also a whole bunch of RISC-V players, catching up year-by-year.

So, first of all, Intel will not stop designing CPUs. In fact, the design side of Intel is the most profitable. If there's anything to worry about here, it's actually their fabs. Second, even if Intel did stop designing new CPUs tomorrow, AMD couldn't rest on its laurels.

look at AMD, they were on the verge of "imploding" and they made it,

You're looking at this through the distorted lens of Hindsight Bias.

AMD was in much worse shape. It was not a given that they would've survived. Maybe, had Intel not suffered so many problems getting newer manufacturing nodes online and production-ready, AMD would be a thing of the past. Many companies in less dire straits have gone under.

 

[bit_user]

I hope 18A is a huge success. TSMC won't have an answer until N2 or A16. Intel is ahead on GaaFET and BSPD. If Intel pulls off 18A, AMD will suddenly be behind on the process tech. They won't have n2 based Epyc until Zen 6 ships in 2026.

Yeah, maybe this time everything will go perfectly!
: D

The thing about TSMC is they didn't have as aggressive a roadmap as Intel, but their execution was much better. They've been regularly hitting their milestones, almost like clockwork. If Intel so much as hits a few snags, its 18A node might end up having mostly to compete with later TSMC nodes. If I'm just going by track records, I know which horse I'd pick.

 

[thestryker]

They've been regularly hitting their milestones, almost like clockwork.

They certainly didn't when they were behind and chasing Intel. Quite frankly they also lucked out that Samsung's 3nm has been so problematic as they were basically able to scrub N3 with minimal loss.

The thing about TSMC is they didn't have as aggressive a roadmap as Intel, but their execution was much better.

As far as recent (N7 and newer) nodes are concerned they've done fantastically which is almost certainly in part to the less aggressive moves.

If Intel so much as hits a few snags, its 18A node might end up having mostly to compete with later TSMC nodes. If I'm just going by track records, I know which horse I'd pick.

It seems fair to say that Intel 3 has been pretty much on track and as expected. There's a lot more riding on 18A though and I think the time frame for execution is the most important part.

 

[bit_user]

It seems fair to say that Intel 3 has been pretty much on track and as expected.

Not sure about that. For the server CPUs they planned to productize on it, it seems on target. However, Intel 4 definitely missed its goals for enabling a suitable successor to Raptor Lake. It could be that Intel 3 also missed some of its goals for being able to target at such products, but we just lack visibility to see that.

 

[thestryker]

It could be that Intel 3 also missed some of its goals for being able to target at such products, but we just lack visibility to see that.

I doubt that's the case if we factor in the mess of Intel 4 (the original plan had GNR on Intel 4).

Not sure about that. For the server CPUs they planned to productize on it, it seems on target. However, Intel 4 definitely missed its goals for enabling a suitable successor to Raptor Lake.

The original 5n4y timeline was basically 12mo between each node and Intel 4 absolutely destroyed that.

I can't help but wonder if they didn't get EUV machines installed as quickly as originally intended (the ASML backlog was really nasty 2020-2022) in concert with the Intel 4 issues. Of course the massive delays caused a lot more node overlap than there otherwise would be so it could also be that simple.

What I find interesting is that Intel 4 was 5-6Q late, Intel 3 about 4Q late (which matches the original time between nodes), 20A whether used or not maybe 1Q late maximum and 18A is still on target (we'll find out as I think skepticism is healthy until there are firm announcements).

 

[jabojabo]

Didn't Broadcom recently complain about how disappointing they found first test drive productions of their chips on Intel's 18A process?

Broadcom supposedly made such comments, but the reports were very non spedific. No one said when the die were made, nor which tests were issues. There are approximately 10 other companies evaluating 18A and no one else has made any negative claims. This was likely an article based on second hand comments by some one who tested 18A die earl

5 nodes in 5 years doesn't count if you outsource one node (20A) to TSMC.

By definition you cannot outsource your own node. A node is a specific set of instructions (a recipe) and rules on how to fabricate chips in a specific way. It's theoretically possible to run a node somewhere else, but they are typically fab specific. You can outsource a design to be fabbed at TSMC using one of their nodes....I guess that's what you're thinking?

 

[jabojabo]

There is so much negative misinformation and click bait floating around about Intel it's dizzying. FACTS: 1)Intel prior CEO prepaid for TSMC N3 capacity. This is the main reason Intel is using TSMC for LUnar Lake and Arrow Lake. Intel is using Intel 3 for Xeon6 and for one external customer. 2) Intel 20A never had plans to be in widespread use. It is/was a stepping stone learning node. I think it had some plans for a limited number of skus on Arrow Lake, but that's it. 3) 18A by all accounts, except one, is doing just fine. Pat even gave some defect/yield numbers recently and those were healthy. 18A is in chips now booting up and running windows. 18A is nearing the finish line and all indications are that it will be just fine. Let's not forget, that 18A has backside power, something dreamed about for over a decade and Intel seems to have this nailed. TSMC N2 will not have backside power. 4) The one negative article about Broadom having issues with 18A is very vague. It didn't specifiy when the chips were made. Maybe they were from early on in the devleopment? No one states this. The Broadcom article says 18A failed some test, but never stated which tests. Finally Broadcom didn't choose to stop interaction with Intel 18A. 5) This article claims to know that TSMC N3 is better than Intel 3. Based on what data? The one and only article I could find trying to make a comparsion said they are about the same. 6) TSMC N2 is not out yet either. TSMC N2 and 18A are both scheduled to begin ramping in 2025. It is not clear which is better. 6) Intel is the one and only company that has so far purchased, installed, and started R&D with an ASML high NA EUV lithograhpy tool. This will be used for Intel 14A in late 2026. TSMC's response, "we don't need high NA". They say this despite ASML claiming it can print feature 1.6x smaller (just NA value ratio). Intel is doing just fine. Pat has gotten them back on track from a technological standpoint. Now they just need to deliver and start shipping products on all these new nodes. This is going to happen. Why all the negativity for Intel. Why do people seem happy about the possibility of Intel failing?

 

[bit_user]

By definition you cannot outsource your own node.

I think there was probably no confusion about that point.

You can outsource a design to be fabbed at TSMC using one of their nodes....I guess that's what you're thinking?

Yeah. In case anyone doesn't know, what happened is that Intel announced it was not going forward with its 20A node. Since the only announced product to be made on 20A was Arrow Lake's compute tile, the presumption is that Intel instead outsourced manufacturing of that chip to TSMC, which makes a lot of sense considering Intel already outsourced manufacturing of Lunar Lake to TSMC (using its own N3B node for the compute tile). Having already done layout of Lion Cove and Skymont on N3B, it was probably rather straight-forward for Intel to layout another chip with more of each, for production on the same node.

For Intel, the main downside is that Arrow Lake is not going to be very profitable. However, outsourcing it to TSMC must've been the least-bad option.

 

[bit_user]

There is so much negative misinformation and click bait floating around about Intel it's dizzying. FACTS: 1)Intel prior CEO prepaid for TSMC N3 capacity.

I've seen this claim, but since you're claiming it's a fact, you should be able to provide a good source on it.

This is the main reason Intel is using TSMC for LUnar Lake

During a Q&A session, I think at a Lunar Lake launch event, the question was asked why Intel used TSMC for the compute tile (whereas they'd previously used the Intel 4 node for Meteor Lake's compute tile). The answer was that the designers were given the choice of what node to use and TSMC's N3 node was the best available option, at the time. So, you're directly contradicting Intel on this.

and Arrow Lake.

Then why did Intel publicly announce that Arrow Lake would use 20A, if they had planned to use TSMC, all along?

Intel 20A never had plans to be in widespread use. It is/was a stepping stone learning node.

That's different than their research nodes, which it sounds like you're trying to claim it was. For instance, Intel made a research node to test out backside power delivery, based on Intel 3, and used it to fab a Crestmont core. That weird hybrid node was never previously announced. It didn't show up on their roadmap and they only mentioned it after the fact.

18A by all accounts, except one, is doing just fine. Pat even gave some defect/yield numbers recently and those were healthy. 18A is in chips now booting up and running windows.

All of these things were said about Intel 4 and Meteor Lake, yet Meteor Lake-S (i.e. the desktop product) ended up getting cancelled and Meteor Lake has been fairly unprofitable for Intel, due to relatively poor yields on the Intel 4 compute tile. I no longer trust what Intel says about their nodes.

Let's not forget, that 18A has backside power, something dreamed about for over a decade and Intel seems to have this nailed.

Yeah, probably first seen on an IMEC fab technology roadmap. That's where a lot of these ideas originate - not from Intel, TSMC, or Samsung. I think it's probably necessary to have someone doing such long-range technology development, so that ASML and other partner companies know what sorts of capabilities future machines need to have.

TSMC N2 will not have backside power.

Presumably, it'll have other features that make up at least some of the difference.

Intel is the one and only company that has so far purchased, installed, and started R&D with an ASML high NA EUV lithograhpy tool.

TSMC has raised cost-related questions around high-NA and has their own roadmap that pursues improvements elsewhere. In my mind, it's not a given that high-NA will confer an unassailable advantage, just like I'm not sure about backside power delivery.

Even if Intel really does pull off a coup and surpass TSMC on key parameters, that's not the end of the story. We have yet to see how Turin (Zen 5-based EPYC) will measure up against Xeon 6, which supposedly will have a node advantage for at least the P-core version.

Intel is doing just fine.

No, they're objectively not. You believe Intel will do just fine, but such predictions have been wrong before. Don't count your chickens before they hatch.

Why all the negativity for Intel. Why do people seem happy about the possibility of Intel failing?

I'm just trying to be realistic. I think Intel has made misleading statements and not been forthcoming with negative information, on numerous occasions under Pat. I am now very skeptical and no longer willing to take their word for anything. From here on out, they're going to prove it to me, first.

And no, I do not want Intel to fail. Not any part of it.