Discussion Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads

[Tigerick]

As Hot Chips 34 starting this week, Intel will unveil technical information of upcoming Meteor Lake (MTL) and Arrow Lake (ARL), new generation platform after Raptor Lake. Both MTL and ARL represent new direction which Intel will move to multiple chiplets and combine as one SoC platform.

MTL also represents new compute tile that based on Intel 4 process which is based on EUV lithography, a first from Intel. Intel expects to ship MTL mobile SoC in 2023.

ARL will come after MTL so Intel should be shipping it in 2024, that is what Intel roadmap is telling us. ARL compute tile will be manufactured by Intel 20A process, a first from Intel to use GAA transistors called RibbonFET.





Intel Core Ultra 100 - Meteor Lake



As mentioned by Tomshardware, TSMC will manufacture the I/O, SoC, and GPU tiles. That means Intel will manufacture only the CPU and Foveros tiles. (Notably, Intel calls the I/O tile an 'I/O Expander,' hence the IOE moniker.)

 

[Exist50]

Considering how Apple Silicon and ARM SoC keep being extended with accelerators and AMD positions AVX-512 and Xilinx's AIE for that purpose it will be really interesting how Intel intends to move forward. Stagnation is not really an option, especially in an environment with strong competition.

Well at least Meteor Lake is supposed to be adding some Movidius IP, so they can use that as their token batch inferencing accelerator. But on the CPU side, seems like they have no choice but to stick with 256b vector solutions for the foreseeable future.

 

[IntelUser2000]

It does not surprise me with Meteorlake. It'll take Arrowlake or successors for the E cores to have AVX-512 instruction set support.

 

[moinmoin]

But on the CPU side, seems like they have no choice but to stick with 256b vector solutions for the foreseeable future.

You mean they are victims of their own ineptness?

 

[mikk]

[Rumor]
Arrow lake-S TSMC N3 series process.
Arrow lake-P INTC 20A series process.

https://twitter.com/x/status/1583277425944690688

You must include your own personal commentary when posting links of any kind.

Daveybrat
AT Moderator

 

[moinmoin]

https://twitter.com/x/status/1583277425944690688

That's about just the compute die/tile, or what parts?

 

[trivik12]

Raichu said its a rumor at this point and he will update if its confirmed. So for now we have to take it with a grain of salt.

https://twitter.com/x/status/1583289828833038336

 

[Exist50]

Raichu said its a rumor at this point and he will update if its confirmed. So for now we have to take it with a grain of salt.

https://twitter.com/x/status/1583289828833038336

I read that more as him saying there's a story behind the split that he'll explain at a later point.

 

[Exist50]

https://twitter.com/x/status/1583277425944690688

Interesting. You'd think it would be the other way around, with 20A being preferred for high performance. Volume-wise, this would imply pretty high volume on both.

 

[Geddagod]

Interesting. You'd think it would be the other way around, with 20A being preferred for high performance. Volume-wise, this would imply pretty high volume on both.

Saw a pretty interesting comment on that tweet about a potential reason, Intel 20A should be coming around 2H 2024, but with TSMC 3nm, they can push a release for 1H 2024. This might be a priority for Intel considering that zen 5 might end up launching 1H 2024 as well.
This theory would also fit in with Raichu and Kopite claiming that MTL would only comprise of the low end stack for 14th gen, while ARL will be high end products released at nearly the same time. MTL-S low end at the end of 2023, ARL-S high end at the start of 2024.
Jim from Adored TV also leaked a while ago that ARL-P was going to use TSMC 3nm, so what might have happened is that they had the entire stack of ARL was on TSMC 3nm originally, then changed ARL mobile skus to Intel 20A for a)diversification of node b)potentially better perf/watt characteristics, all of which was enabled by pushing back ARL mobile skus a year from their original released date (at least according to Jim) while instead launching ARL-S in its place.
But, this is just speculation

 

[mikk]

Saw a pretty interesting comment on that tweet about a potential reason, Intel 20A should be coming around 2H 2024, but with TSMC 3nm, they can push a release for 1H 2024. This might be a priority for Intel considering that zen 5 might end up launching 1H 2024 as well.
This theory would also fit in with Raichu and Kopite claiming that MTL would only comprise of the low end stack for 14th gen, while ARL will be high end products released at nearly the same time. MTL-S low end at the end of 2023, ARL-S high end at the start of 2024.
Jim from Adored TV also leaked a while ago that ARL-P was going to use TSMC 3nm, so what might have happened is that they had the entire stack of ARL was on TSMC 3nm originally, then changed ARL mobile skus to Intel 20A for a)diversification of node b)potentially better perf/watt characteristics, all of which was enabled by pushing back ARL mobile skus a year from their original released date (at least according to Jim) while instead launching ARL-S in its place.
But, this is just speculation

Exactly, also Raichu said or implied MTL-S and ARL-S are coming close together which makes sense considering the claims that MTL-S will be lowend-midrange and ARL-S highend on the same plattform. Launching a new plattform which is limited to lowend-midrange for 9-12 months would be really bad. With TSMC 3nm they can bring ARL-S in Q1 or Q2 2024 which isn't realistic on Intel 20A. Another issue could be volume. The entire mobile stack and highend Desktop all on 20A might be difficult to achieve from day one.

 

[Exist50]

Saw a pretty interesting comment on that tweet about a potential reason, Intel 20A should be coming around 2H 2024, but with TSMC 3nm, they can push a release for 1H 2024. This might be a priority for Intel considering that zen 5 might end up launching 1H 2024 as well.
This theory would also fit in with Raichu and Kopite claiming that MTL would only comprise of the low end stack for 14th gen, while ARL will be high end products released at nearly the same time. MTL-S low end at the end of 2023, ARL-S high end at the start of 2024.
Jim from Adored TV also leaked a while ago that ARL-P was going to use TSMC 3nm, so what might have happened is that they had the entire stack of ARL was on TSMC 3nm originally, then changed ARL mobile skus to Intel 20A for a)diversification of node b)potentially better perf/watt characteristics, all of which was enabled by pushing back ARL mobile skus a year from their original released date (at least according to Jim) while instead launching ARL-S in its place.
But, this is just speculation

The timeline seems plausible, but then why prioritize desktop? Surely they'll be desperate for a competitor for Zen 5, but desperate enough to prioritize the historically less-important desktop market over mobile? And on top of that, we previously had some evidence for a flagship ARL mobile halo product planned for this year.



That's clearly not happening for one reason or another, but was it cancelled altogether? Shifted out an entire year? And for what particular reason? Conflict with MTL? Lion Cove delays? Skymont delays?

Launching a new plattform which is limited to lowend-midrange for 9-12 months would be really bad.

I'm not seeing any good scenarios regardless. If they're close together, then that implies MTL is H2'23 at best, which implies pretty severe delays beyond Intel 4. And either way, they really won't have anything for enthusiast desktop till whenever Arrow Lake happens to show up, and who knows when that'll really be.

 

[mikk]

I'm not seeing any good scenarios regardless. If they're close together, then that implies MTL is H2'23 at best, which implies pretty severe delays beyond Intel 4. And either way, they really won't have anything for enthusiast desktop till whenever Arrow Lake happens to show up, and who knows when that'll really be.

MTL-P is H2 2023 by the looks of it, MTL-S is later than this, maybe H1 2024. Remember Zen 5 is coming not before 2024, Intel is not in a hurry for MTL-S in 2023. 20A is production ready in H1 2024 they say, means no CPUs in stores until H2 2024 and likely late 2024. They can launch both MTL-S and ARL-S during H1 2024. This is not possible on Intel 20A.

 

[Geddagod]

The timeline seems plausible, but then why prioritize desktop? Surely they'll be desperate for a competitor for Zen 5, but desperate enough to prioritize the historically less-important desktop market over mobile? And on top of that, we previously had some evidence for a flagship ARL mobile halo product planned for this year.

View attachment 69597

That's clearly not happening for one reason or another, but was it cancelled altogether? Shifted out an entire year? And for what particular reason? Conflict with MTL? Lion Cove delays? Skymont delays?

I'm not seeing any good scenarios regardless. If they're close together, then that implies MTL is H2'23 at best, which implies pretty severe delays beyond Intel 4. And either way, they really won't have anything for enthusiast desktop till whenever Arrow Lake happens to show up, and who knows when that'll really be.

For desktop, before ARL, all they would have is MTL-S, which is supposed to be only focusing on the low end of desktop. Maybe it's because they won't have high core counts, maybe it's because they can't clock as well, but whatever it is, for a high performance desktop sku, Intel won't have much. Against a zen 5 desktop sku, I doubt MTL would be able to compete, especially since we don't even know if MTL goes up to 8+16 for sure! For mobile, before ARL, they would have MTL mobile skus, and even if they don't win on sheer performance vs a 16 core dragon range-x sku, they should be able to, at the very least, be competitive in efficiency/battery life. And in laptops, that is a much better position to be in, than in desktop where perf matters a lot more.
We also don't know the launch schedule for Zen 5 either, but I would expect mobile to come after desktop skus, so if mobile comes 1-2 quarters after desktop for zen 5, then Intel might be taking advantage of that schedule to release ARL mobile on a better node (Intel 20A) 2h 2024.
Or maybe they just aren't confident in their mobile competitiveness of ARL vs Zen 5 without a node advantage of using 20A vs TSMC 3nm.
I see a lot of potential reasons why Intel might to do this. Who knows, maybe it is all baloney, but if it's true, I can see some of the reasoning for it at least.

For the ARL-P being scheduled for this year, isn't that roadmap showing that ARL-P is scheduled for the end of 2023? I'm guessing ARL got delayed for the same reason a lot of Intel stuff got delayed, rolling back schedules from the original delay in products. For example, since MTL mobile/desktop got pushed into 2H 2023, or maybe a limited release 1H 2023, I think the earliest ARL could release is 1H 2024 (Ik Rocket Lake released a lot quicker but that being on the same node on a backported architecture seems like an exception). And that sounds like what Raichu thinks Intel is pushing to do.

 

[Exist50]

Remember Zen 5 is coming not before 2024, Intel is not in a hurry for MTL-S in 2023.

Intel needs Meteor Lake to compete with Zen 4, not Zen 5. They're going to get it out as soon as possible. Arrow Lake is what they'll need to compete with Zen 5.

They can launch both MTL-S and ARL-S during H1 2024. This is not possible on Intel 20A.

Ok, but that leaves the same question of why desktop, and not mobile? The MTL/ARL split could theoretically apply to both. Really, if ARL actually arrived at the same time as MTL, the only reason the latter would have to exist at all is to avoid the cost/fab implications of having everything on N3 or 20A.

 

[A///]

The timeline seems plausible, but then why prioritize desktop? Surely they'll be desperate for a competitor for Zen 5, but desperate enough to prioritize the historically less-important desktop market over mobile? And on top of that, we previously had some evidence for a flagship ARL mobile halo product planned for this year.

Volume. If Intel can offload to TSMC and have them produce at volume while they also produced volume it would clear them at home to make more while TSMC did their thing. It's quite genius if you think about it. Intel may have realized they may not have capacity then to produce both at home.

Intel knows AMD won't stand still. The KS processor is going to be hot but powerful, and X3D may take the gaming crown. Intel has a lot to lose still by standing still. They need to get both desktop and mobile out. AMD's upcoming mobile solutions are going to be something fierce and they need to compete. I reckon Intel will drop TSMC later when their capacity goes up.

This also lets Intel focus on getting their foundry services set up. Gelsinger legitimately wants people to use Intel for their production needs because it's domestic and not foreign like TSMC. He's even mentioned that he'd love to have AMD contract them. There is a mutual benefit there I'm sure.

 

[Geddagod]

Volume. If Intel can offload to TSMC and have them produce at volume while they also produced volume it would clear them at home to make more while TSMC did their thing. It's quite genius if you think about it. Intel may have realized they may not have capacity then to produce both at home.

Intel knows AMD won't stand still. The KS processor is going to be hot but powerful, and X3D may take the gaming crown. Intel has a lot to lose still by standing still. They need to get both desktop and mobile out. AMD's upcoming mobile solutions are going to be something fierce and they need to compete. I reckon Intel will drop TSMC later when their capacity goes up.

This also lets Intel focus on getting their foundry services set up. Gelsinger legitimately wants people to use Intel for their production needs because it's domestic and not foreign like TSMC. He's even mentioned that he'd love to have AMD contract them. There is a mutual benefit there I'm sure.

The question isn't why Intel will use TSMC as a foundry, it's why use TSMC as a foundry for desktop instead of mobile ARL.
I also think Intel will drop TSMC 3nm for ARL-S too, but not just for volume reasons. If Intel 20A is indeed better than TSMC 3nm, which from what I hear it seems to be, then they can easily just do a refresh of ARL-S onto Intel 20A or maybe Intel 18A (since 18A seems to be the nodelet from Intel 20A with the same design rules). This would also make further sense since ARL mobile is already supposed to be on Intel 20A so it's not even like they need to redesign and port the entire architecture for a generational refresh.

 

[A///]

The question isn't why Intel will use TSMC as a foundry, it's why use TSMC as a foundry for desktop instead of mobile ARL.
I also think Intel will drop TSMC 3nm for ARL-S too, but not just for volume reasons. If Intel 20A is indeed better than TSMC 3nm, which from what I hear it seems to be, then they can easily just do a refresh of ARL-S onto Intel 20A or maybe Intel 18A (since 18A seems to be the nodelet from Intel 20A with the same design rules). This would also make further sense since ARL mobile is already supposed to be on Intel 20A so it's not even like they need to redesign and port the entire architecture for a generational refresh.

One theory is Intel did the math and figured out that having TSMC make their core dies is cheaper than doing it at home and the process may be efficient. They would have node parity with AMD at the same foundry. If Intel's node is more power efficient than TSMC 3nm then it will benefit mobile a lot more than desktop. Intel's suffered a lot in desktop and datacenter. One good generation isn't going to cut it. They don't have a real answer for datacenter. They still reign supreme in mobile. I expect PCIe6 to be around by then. We won't see DDR6 until 2028-2029.

I'm not worried about 3nm volume. If TSMC is only making the core dies and Intel is doing the rest at home on their older processes it's up to Intel to determine how much volume they ship out. I would be worried about volume if Intel were to producing all at home. It wasn't too long ago Intel was having issues with expensive nodes not producing enough volume. I don't think Gelsinger wants a repeat of that.

 

[Exist50]

One theory is Intel did the math and figured out that having TSMC make their core dies is cheaper than doing it at home

There's no chance that N3 is cheaper than any of Intel's internal nodes.

 

[mikk]

Intel needs Meteor Lake to compete with Zen 4, not Zen 5. They're going to get it out as soon as possible. Arrow Lake is what they'll need to compete with Zen 5.

Raptor Lake-S is more than good enough against Zen 4. Actually Zen 4 looks quite bad in the mid tier range, I mean 13600K beats 7700x in many cases and it's cheaper. Against Zen 5 they need something new, I doubt MTL-S is enough. Maybe in the lower end but certainly not for the higher end.

 

[Geddagod]

One theory is Intel did the math and figured out that having TSMC make their core dies is cheaper than doing it at home and the process may be efficient. They would have node parity with AMD at the same foundry. If Intel's node is more power efficient than TSMC 3nm then it will benefit mobile a lot more than desktop. Intel's suffered a lot in desktop and datacenter. One good generation isn't going to cut it. They don't have a real answer for datacenter. They still reign supreme in mobile. I expect PCIe6 to be around by then. We won't see DDR6 until 2028-2029.

I'm not worried about 3nm volume. If TSMC is only making the core dies and Intel is doing the rest at home on their older processes it's up to Intel to determine how much volume they ship out. I would be worried about volume if Intel were to producing all at home. It wasn't too long ago Intel was having issues with expensive nodes not producing enough volume. I don't think Gelsinger wants a repeat of that.

Like Exist 50 said, I highly doubt TSMC 3nm would be cheaper for Intel than in house. They also don't need TSMC 3nm to be competitive with AMD node on node, even if Intel 3 isn't as good as TSMC 3nm, then their Intel 20A at the very least should be. And Intel 20A should be ready for mass production by 2H 2024, and considering Zen5 is rumored to be 1H 2024, it doesn't make much sense to move ARL to TSMC 3nm unless they wanted ARL out immediately to zen 5. But why do they want ARL-S out immediately, rather than ARL-P, the more important product?
If it's a volume thing, it makes more sense to move ARL mobile to TSMC rather than ARL-S.
Edit: Samsung announced ddr6 could be available by 2024, but I'm willing to bet we won't actually see it until 2025/2026. I don't have the data for this, but can someone check how long ddr4/ddr3/ddr2 lasted relative to each other? I think ddr4 lasted exceptionally long.

 

[Geddagod]

Raptor Lake-S is more than good enough against Zen 4. Actually Zen 4 looks quite bad in the mid tier range, I mean 13600K beats 7700x in many cases and it's cheaper. Against Zen 5 they need something new, I doubt MTL-S is enough. Maybe in the lower end but certainly not for the higher end.

It's not RPL-S that can't compete, problem could be Raptor Lake mobile. While Intel does seem to have some tricks up their sleeve with DLVR? I think it was, and probably some other minor improvements, I still expect a 5nm Zen 4 to be vastly more efficient. And for the 16 core Zen 4 mobile cpu to blow RPL out of the water in perf/watt and even just perf because you can't clock RPL too high in laptops because of heat.
I still kinda disagree with Exist 50 statement though, because I doubt MTL brings much higher performance than RPL-S. Slight IPC bump? Sure. Frequency increase from new node? Ok. But MTL would also be plagued by new problems- chiplets are inherently less efficiency than monolithic, because of sending energy between tiles. I expect MTL to help, but also be pretty limited in availability.

 

[moinmoin]

Raptor Lake-S is more than good enough against Zen 4.

Except that Raptor Lake-S is only available as top end desktop chips for inexplicable reasons. Different older chips will have to make do against Zen 4 in mobile and servers.

 

[Exist50]

Raptor Lake-S is more than good enough against Zen 4.

RPL-S is surprisingly competitive against Zen 4 in desktops, but it's going to get crushed in mobile. MTL should help there, but day late and a dollar short. If Arrow Lake is actually an H1'24 product, then that should be very significant for having not only a process advantage, but also an actual architectural upgrade over Raptor Lake. Which is why I'm confused that they're supposedly spending this advantage on one of the least important markets. But maybe they just can't afford to not have a competitive offering for that long.

Edit: Samsung announced ddr6 could be available by 2024, but I'm willing to bet we won't actually see it until 2025/2026.

I think even '25/'26 might be optimistic. Probably the very earliest we could expect it would be Nova Lake from Intel or whatever Zen 7 will be called on the AMD side.

 

[mikk]

Except that Raptor Lake-S is only available as top end desktop chips for inexplicable reasons. Different older chips will have to make do against Zen 4 in mobile and servers.

We are talking about desktop not mobile. Exist50 basically claimed Raptor Lake-S is not good enough and they need MTL-S against Zen 4 which I and the reviews disagree with. Once again Intel is not in a hurry to bring MTL-S in 2023 because Zen 5 won't come before 2024. If anything AMD needs a price cut to compete with Intels offerings. There is lots of software and preparation work going on for mobile MTL right now and nothing for MTL-S, it's more than obvious Intel favours mobile MTL. Historically this is not a surprise. 14nm, 10nm....and possibly 20A Intel favours mobile on a new process.

 

[moinmoin]

@mikk Even if limiting to desktop you seem to ignore the existence of later V-Cache variants of Zen 4.