Discussion Intel current and future Lakes & Rapids thread

[Exist50]

The smallest chip shown is Foveros. The larger ones with more P-cores and/or more graphics, will for sure use Foveros Omni to maximize reuse.

"For sure", no... It's wrong to assume that Foveros Omni is ready for Meteor Lake, especially since MTL was originally hoped to be a 2022 product. Might have been nice to have, but Intel would have highlighted it if MTL were going to use it. I'll be generous and say "Semianalysis" is just being overly optimistic.

I believe @Exist50 though when he says Meteorlake will have little desktop presence. So I'd say no on the 8+16. This gives them an opportunity to speed up Arrowlake-S launch as well.

I have to say, I'm very curious what they do with Arrow Lake. From an older leak, we know they were planning to use N3 for ARL Halo compute, so the question is, what about 20A? I think the most likely scenario is use N3 to begin with and refresh or dual source on 20A, but schedule will be quite interesting. If they're struggling to get both MTL and ARL out, I really doubt adding a "proper" MTL desktop is in the cards, and I haven't heard of one, but it does depend on the exact reason.

So it does seem like your RPL-R scenario is quite plausible (even the most likely), but man would that be disappointing...

With "10-12%" gain.

Oh, yeah, almost forgot that. Tough to say what he meant though. Hopefully, given the context, at least iso-process, but also iso-power? Iso-area? If one or both, worth the risk.

 

[ashFTW]

"For sure", no... It's wrong to assume that Foveros Omni is ready for Meteor Lake, especially since MTL was originally hoped to be a 2022 product. Might have been nice to have, but Intel would have highlighted it if MTL were going to use it. I'll be generous and say "Semianalysis" is just being overly optimistic.

I believe that MTL products larger than 2+8+64EU, are the test vehicle for Foveros Omni. MTL has gotten delayed due to process, but Omni may be in reasonable shape already.


Edit: Wikichip says MTL will use Foveros Omni.

”Meteor Lake is Intel’s upcoming Intel 4-based product slated for 2023 for the client market. Meteor lake will make sure use of Foveros Omni with a 36-micron micro bump pitch. The processor will make sure use of multiple dies from different process nodes and will span from 5 W all the way to 125 W.”

 

[Exist50]

MTL has gotten delayed due to process

Design/validation, to the best of my understanding. Though process churn doesn't help.

 

[ashFTW]

Design/validation, to the best of my understanding. Though process churn doesn't help.

I don’t doubt the design/validation challenges; it’s Intel’s first (not counting Ponte Vecchio) real (not counting Lakefield) disaggregated chip design using Foveros.

With Foveros Omni, during packaging, Copper columns are used to bypass the base die to directly supply power and signals from the package to the top die. I expect these steps to not affect the design of the tiles themselves, beyond what Foveros itself requires. Foveros Omni test wafers were shown a year ago (source WikiChip) at Intel Accelerated ‘21.

 

[lobz]

What is new? Arrow Lake on Lion Cove+Skymont, 20A and 8+32 isn't new, although he confirmed the older leaks. Meteor Lake using Redwood Cove+Crestmont isn't new either, IPC estimates are too vague at the moment. To me it isn't new but of course not every is aware of all leaks. LGA socket number is definitely new and he even posted a picture which once again shows he has real sources (when it comes to Intel). I haven't seen the screenshot from discord, I cannot comment on this.

The real(ly concerning) news for me in this vid are the confirmation of even further delays. Especially when it comes to raptor lake.

 

[lobz]

Maybe I'll regret making such strong claims, but I'm sticking to my guns for now. I think they're justified based on the available information.

Claims? Pretty much all you wrote was bashing.

 

[lobz]

Same goes for you, my friend!

That has to going to be a very interesting launch then, if basically everyone's going to be disappointed, pro or contra

 

[mikk]

With "10-12%" gain.

This is not what Gelsinger said, he said 10-plus percent which can mean anything above 10 percent. We will see if this is as accurate as the up to double digits performance claim for Raptor Lake where people assumed it must be 10 percent.

Meteor Lake and Arrow Lake is on LGA-1851: https://videocardz.com/newz/intel-l...el-meteor-lake-and-arrow-lake-has-been-leaked

 

[uzzi38]

The smallest chip shown is Foveros. The larger ones with more P-cores and/or more graphics, will for sure use Foveros Omni to maximize reuse.

Ian Cutress asked about the Foveros used for MTL and was told very clearly it is 36um Foveros, no mention of Omni at all.

https://twitter.com/x/status/1429594452960681994

 

[IntelUser2000]

This is not what Gelsinger said, he said 10-plus percent which can mean anything above 10 percent. We will see if this is as accurate as the up to double digits performance claim for Raptor Lake where people assumed it must be 10 percent.

Try again.

Trying to counter me every single time again eh?

He did say 10-plus but he also said 10-12%.

 

[IntelUser2000]

Edit: Wikichip says MTL will use Foveros Omni.

”Meteor Lake is Intel’s upcoming Intel 4-based product slated for 2023 for the client market. Meteor lake will make sure use of Foveros Omni with a 36-micron micro bump pitch. The processor will make sure use of multiple dies from different process nodes and will span from 5 W all the way to 125 W.”

So they got it wrong. While Meteorlake does use second generation Foveros, Omni is farther away.

Anandtech article also says the same thing, with Omni being 3rd generation: https://www.anandtech.com/show/1682...-3nm-20a-18a-packaging-foundry-emib-foveros/4

On a side note, I don't know why articles like from Semianalysis makes a big deal about package size and mobility designs, especially for laptops.

They already make really compact designs from OneNetbook and such. Since Haswell, the chipset has been part of the package. Cutting that in half won't change things too much. So you have the board that's much much larger, but reducing the size of the chip that's less than 5% of the area will change things? Really?

Maybe in Tablet designs, but they got other things to worry about like power use and Jasper Lake with not so small package also has Tablet designs. This focus on package size is like looking at trees instead of the forest.

The bigger impact to overall design size is TDP, because heatsinks are quite large, and efficiency, so the battery doesn't have to be humongous.

"Let's cut the package size in half, that's going to be awesome in making tiny laptops!"
"Uhh, but what about the 125W TDP? Surely that factors in device size a tad?"

 

[nicalandia]

So Desktop MTL/Arrow lake Socket will be about the same size as ADL/RTL

 

[lobz]

So they got it wrong. While Meteorlake does use second generation Foveros, Omni is farther away.

Anandtech article also says the same thing, with Omni being 3rd generation: https://www.anandtech.com/show/1682...-3nm-20a-18a-packaging-foundry-emib-foveros/4

On a side note, I don't know why articles like from Semianalysis makes a big deal about package size and mobility designs, especially for laptops.

They already make really compact designs from OneNetbook and such. Since Haswell, the chipset has been part of the package. Cutting that in half won't change things too much. So you have the board that's much much larger, but reducing the size of the chip that's less than 5% of the area will change things? Really?

Maybe in Tablet designs, but they got other things to worry about like power use and Jasper Lake with not so small package also has Tablet designs. This focus on package size is like looking at trees instead of the forest.

The bigger impact to overall design size is TDP, because heatsinks are quite large, and efficiency, so the battery doesn't have to be humongous.

"Let's cut the package size in half, that's going to be awesome in making tiny laptops!"
"Uhh, but what about the 125W TDP? Surely that factors in device size a tad?"

Not only that, larger raw square mm footprint could even be far more desirable in mobile, provided it's thin enough and handles hotspots well.

 

[igor_kavinski]

News release - A miniature microfluidics heat sink For high-performance chip cooling (imec-int.com)

Silicon as a material is a relatively good heat conductor. The use of small, parallel, high-aspect-ratio silicon microchannel structures of 32µm wide and more than 260µm deep in imec’s chip cooler further increases the convective heat transfer surface area and the heat transfer coefficient, enabling high heat flux removal. This makes it possible to dissipate power of more than 600W/cm2 while keeping the component temperature below 100°C.

This will become a necessity in the Angstrom era.

 

[mikk]

According to Raichu and kopite7kimi MTL-S will be a mid to lowend Mainstream series for desktop while ARL-S is coming for the highend. I believe them. MTL-S 6+8 maybe the maximum?

https://twitter.com/x/status/1533759998186393600

https://twitter.com/x/status/1533783931044376576

 

[nicalandia]

According to Raichu and kopite7kimi MTL-S will be a mid to lowend Mainstream series for desktop while ARL-S is coming for the highend. I believe them. MTL-S 6+8 maybe the maximum?

https://twitter.com/x/status/1533759998186393600

https://twitter.com/x/status/1533783931044376576

So Raptor Lake will be left alone to fend off High Core Count Zen4/Zen5 AM5 CPUS until arrow lake on 2024?

 

[mikk]

So Raptor Lake will be left alone to fend off High Core Count Zen4/Zen5 AM5 CPUS until arrow lake on 2024?

Yes seems like this is the case, although the H2 2024 timeframe from MLID might be not accurate. Raichu says MTL-S and ARL-S will come almost at the same time. Or it means MTL-S is more towards H2 2024 as well. And then Lunar Lake might be only for low power, the next bigger one is apparently Panther Lake.

 

[moinmoin]

So Desktop MTL/Arrow lake Socket will be about the same size as ADL/RTL

View attachment 62715

So the only difference appears to be an increased thickness of 0.09-0.1mm.

View attachment 62716

So Raptor Lake will be left alone to fend off High Core Count Zen4/Zen5 AM5 CPUS until arrow lake on 2024?

Would be surprising but not the first time for Intel to take its sweet time for continuing to scale up. If AMD projected that it makes sense to continue "only" with 16 cores max on AM5 for the time being.

 

[nicalandia]

Yes seems like this is the case, although the H2 2024 timeframe from MLID might be not accurate. Raichu says MTL-S and ARL-S will come almost at the same time. Or it means MTL-S is more towards H2 2024 as well. And then Lunar Lake might be only for low power, the next bigger one is apparently Panther Lake.

No chance of ARL-S coming in 2023.

This does not look very good because intel will have to fight AMD Bergamo/Turin with 10nm High End Desktop/UHEDT/Server Raptor Cove core until 2024 where Granite Rapids is scheduled to be released with a new core on Intel 3

 

[mikk]

No chance of ARL-S coming in 2023.

This does not look very good because intel will have to fight AMD Bergamo/Turin with High End Desktop/UHEDT/Server Raptor Cove until 2024 where Granite Rapids is scheduled to be released

Assuming Zen 5 is ready in H2 2023 and ARL-S 8+32 in H2 2024 then it could be a difficult year for the desktop. Mainstream part of Raptor Lake seems to come early 2023 at CES and therefore MTL-S lowend-mainstream is coming more likely at CES 2024 in the best case I think. Or maybe later like Q2 2024 and then Q3 2024 ARL-S. This is almost the same time to me.

 

[Exist50]

I don’t doubt the design/validation challenges; it’s Intel’s first (not counting Ponte Vecchio) real (not counting Lakefield) disaggregated chip design using Foveros.

From what I've heard, it's not so much Foveros or even disaggregation itself, but rather mostly SoC-level fabric changes, compounded by those other issues. Apparently there are some huge uncore changes, and at the same time, they lost a lot of key engineers in that area, so getting that sorted out has been a mess, and likely results in a sub-optimal SoC.

Actually, I heard from a former Lakefield engineer that they were all kinda surprised how well Foveros worked in and of itself.

This is not what Gelsinger said, he said 10-plus percent which can mean anything above 10 percent. We will see if this is as accurate as the up to double digits performance claim for Raptor Lake where people assumed it must be 10 percent.

mikk, as pointed out, Gelsinger said 10-12%. The most optimistic interpretation would be within the same process, power, and area envelope as Redwood Cove, which would be a really impressive achievement. That doesn't necessary tell us anything about single thread, but I'd guess closer to the 10-20% IPC ballpark than some of the numbers being thrown around.

Meteor Lake and Arrow Lake is on LGA-1851: https://videocardz.com/newz/intel-l...el-meteor-lake-and-arrow-lake-has-been-leaked

Ok, I think this makes a lot more sense than that LGA-2551 rumor.

They already make really compact designs from OneNetbook and such. Since Haswell, the chipset has been part of the package. Cutting that in half won't change things too much. So you have the board that's much much larger, but reducing the size of the chip that's less than 5% of the area will change things? Really?

I mean, package size does matter to some degree, especially with the ultra-mobile designs that use PMICs for everything. But advanced packaging does pretty much nothing for them in that regard, so I'm not even sure why it's being mentioned. It would be cool if they could shift power delivery to the base tile, but the claim that's happening for MTL is just another of their wildly off-base guesses. Thermals would be a nightmare to solve too.

Not only that, larger raw square mm footprint could even be far more desirable in mobile, provided it's thin enough and handles hotspots well.

Definitely not. Cost and board space are far more dominant concerns there than cooling.

According to Raichu and kopite7kimi MTL-S will be a mid to lowend Mainstream series for desktop while ARL-S is coming for the highend. I believe them. MTL-S 6+8 maybe the maximum?

Ok, now this sounds right. Reusing the 6+8 mobile die with a new desktop SoC die (and IO die?), and following up with Arrow Lake for the higher end market?

Of course, that means Intel will be competing with reheated Alder Lake all the way till some indeterminate time in 2024, which is a terrible position indeed. Even when Arrow Lake comes, I worry it'll be dragged down by the Meteor Lake holdovers. Maybe 2025 for a proper return to the enthusiast desktop market? Ugh...

 

[nicalandia]

Of course, that means Intel will be competing with reheated Alder Lake all the way till some indeterminate time in 2024, which is a terrible position indeed. Even when Arrow Lake comes, I worry it'll be dragged down by the Meteor Lake holdovers. Maybe 2025 for a proper return to the enthusiast desktop market? Ugh...

Assuming Zen 5 is ready in H2 2023 and ARL-S 8+32 in H2 2024 then it could be a difficult year for the desktop.

Difficult for Desktop and Plain Awful for Data Center/Severs....

 

[JasonLD]

Assuming Zen 5 is ready in H2 2023 and ARL-S 8+32 in H2 2024 then it could be a difficult year for the desktop. Mainstream part of Raptor Lake seems to come early 2023 at CES and therefore MTL-S lowend-mainstream is coming more likely at CES 2024 in the best case I think. Or maybe later like Q2 2024 and then Q3 2024 ARL-S. This is almost the same time to me.

if Zen 5 uses TSMC 5nm variant, then it could be ready by late 2023. If 3nm, I would expect 2024. 2023 could be without any new desktop high end CPU from either company.

 

[msj10]

I hope intel has a backup plan for arrow lake if 20A hits any trouble. maybe they can backport Lion cove and Skymont to intel 3 if needed.

 

[nicalandia]

if Zen 5 uses TSMC 5nm variant, then it could be ready by late 2023. If 3nm, I would expect 2024. 2023 could be without any new desktop high end CPU from either company.

No Chance AMD is using N3 from TSCM before Apple, they are schedule to mass produce N3 late 2023 and nowing AMD Process node and release cycle (Zen/Zen+ using 14nm/12nm, Zen2/Zen3 using N7, we can expect Zen4/Zen5 to be on N5)