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.



Comparison of upcoming Intel's U-series CPU: Core Ultra 100U, Lunar Lake and Panther Lake

Model	Code-Name	Date	TDP	Node	Tiles	Main Tile	CPU	LP E-Core	LLC	GPU	Xe-cores
Core Ultra 100U	Meteor Lake	Q4 2023	15 - 57 W	Intel 4 + N5 + N6	4	tCPU	2P + 8E	2	12 MB	Intel Graphics	4
?	Lunar Lake	Q4 2024	17 - 30 W	N3B + N6	2	CPU + GPU & IMC	4P + 4E	0	8 MB	Arc	8
?	Panther Lake	Q1 2026 ?	?	Intel 18A + N3E	3	CPU + MC	4P + 8E	4	?	Arc	12

Comparison of die size of Each Tile of Meteor Lake, Arrow Lake, Lunar Lake and Panther Lake

	Meteor Lake	Arrow Lake (20A)	Arrow Lake (N3B)	Arrow Lake Refresh (N3B)	Lunar Lake	Panther Lake
Platform	Mobile H/U Only	Desktop Only	Desktop & Mobile H&HX	Desktop Only	Mobile U Only	Mobile H
Process Node	Intel 4	Intel 20A	TSMC N3B	TSMC N3B	TSMC N3B	Intel 18A
Date	Q4 2023	Q1 2025 ?	Desktop-Q4-2024 H&HX-Q1-2025	Q4 2025 ?	Q4 2024	Q1 2026 ?
Full Die	6P + 8P	6P + 8E ?	8P + 16E	8P + 32E	4P + 4E	4P + 8E
LLC	24 MB	24 MB ?	36 MB ?	?	8 MB	?
tCPU	66.48
tGPU	44.45
SoC	96.77
IOE	44.45
Total	252.15

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.)

 

[H433x0n]

No, even if they made a reticle limit DG2-1024 at 450W they'd still be behind the 7900 XTX in raytracing performance. So-called architectural advantage does not exist if you can't build bigger. It doesn't exist so they're not number 2. Same thing with AMD saying one of the advantages of their multi-chip architecture is that they could build a supermassive 400mm + 8 MCD RDNA3 chip to match the 4090. OK cool, but they didn't so it's whiteboard cope & dreams and they stay #2.

To summarize: if your raster performance is 0.5x RDNA3 per area on iso process it doesn't matter that you have -23% RT performance and they have -45% you're still 0.7x RT performance per area on iso process. Their raster is just really, really bad.

Let's setup a hypothetical:

RDNA4 improves RT performance drastically and enabling RT incurs a 30% performance hit. However, the 7900 XTX still outperforms it because it has a higher raster perf baseline. Would you consider RDNA4 RT perf to be a regression in that case? If you do, I can understand it, but I would disagree.

To me in that example scenario, I would still say RDNA4 has improved RT performance.

 

[Khato]

With respect to Intel's graphics, the better comparison at current is between Meteor Lake and Phoenix/Hawk Point. GPU die size is comparable despite the process node advantage for AMD. Intel may well be a bit slower and less efficient, but it's close. It'll be interesting to see what kind of improvements Intel brings with Arrow Lake and Lunar Lake compared to AMD's Strix Point. The additional CUs will almost certainly best the minor gains on Arrow Lake, but how it fares against the larger Xe2 based Lunar Lake will be most interesting.

 

[gdansk]

To me in that example scenario, I would still say RDNA4 has improved RT performance.

It has improved RT performance but would have done nothing to become #1 in raytracing. Which was my problem with Igor's claim of Intel being #2.

 

[gdansk]

With respect to Intel's graphics, the better comparison at current is between Meteor Lake and Phoenix/Hawk Point.

While it would be most relevant to this thread we can't normalize for the rest of the test configuration.

 

[Khato]

While it would be most relevant to this thread we can't normalize for the rest of the test configuration.

I take it you're referring to memory bandwidth? Or power consumption?

Regardless, the interesting point to keep in mind with Intel's graphics is that DG2 was effectively Intel's first serious attempt at a large, high performance part. Which was pretty obvious with the resulting poor PPA. Meteor Lake fixed some of the issues with the design, Xe2 will fix even more. I'm still surprised that Meteor Lake is pretty much on par with RDNA3.

 

[gdansk]

I take it you're referring to memory bandwidth? Or power consumption?

Regardless, the interesting point to keep in mind with Intel's graphics is that DG2 was effectively Intel's first serious attempt at a large, high performance part. Which was pretty obvious with the resulting poor PPA. Meteor Lake fixed some of the issues with the design, Xe2 will fix even more. I'm still surprised that Meteor Lake is pretty much on par with RDNA3.

Yes to both. Even CPU performance differences and sustained cooling.

MTL seems much closer but since they're both in the realm of "RT performance is useless" it doesn't support Igor's claim either. But it bodes well for BM.

 

[DrMrLordX]

RDNA4 improves RT performance drastically and enabling RT incurs a 30% performance hit. However, the 7900 XTX still outperforms it because it has a higher raster perf baseline. Would you consider RDNA4 RT perf to be a regression in that case?

Yes. It would be a total regression in performance by every metric.

 

[DavidC1]

Compared to MTL-H's iGPU, below is the major changes in ARL's iGPU:

• Doubled L2 cache.
• Doubled RT throughput.
• XMX cores added.

Are you going to address my questions I asked earlier Tigerick?

To me in that example scenario, I would still say RDNA4 has improved RT performance.

Yes, but that would be a very unbalanced architecture. If the die sizes are same, it means this hypothetical RDNA4 has dedicated too much to improving RT performance when it has to be balanced for all workloads, including raster.

It's equal to if Zen 5/Lion Cove reduces CPU performance by 20% but improves AI performance by 20x. Yay?

 

[SiliconFly]

...nothing special or interesting.

Latest advanced 3D packaging technologies like foveros are very interesting when compared to old and outdated substrate based 2.5D tech like in Zen5.

It's IOSF. From 2012.

No. MTL has a brand new NOC.

From AnandTech itself:
"The other two tiles Intel implements within Meteor Lake are the SoC tile which acts as the central hub through the embedded NOC. This is the first time Intel has used a Network-on-Chip (NOC) on their client processors,... ...The NOC itself is directly connected to the graphics tile, compute tile, and other components within the SoC."

Same applies to ARL and beyond.

 

[DavidC1]

Latest advanced 3D packaging technologies like foveros are very interesting when compared to old and outdated substrate based 2.5D tech like in Zen5.

Ok, but Meteorlake sometimes underperforms the predecessor at the same clock, while Zen series was faster in every way, and competitive with monolithic Raptorlake.

You say fancy things, you dress it up fancy, but it ends up being plane jane. How many people outside of Intel fans will see it otherwise?

They used to lead in DDR implementation, memory controller performance, and caches. Ah, but it's just the fault of the "process" right? RIGHT?

 

[SiliconFly]

Ok, but Meteorlake sometimes underperforms the predecessor at the same clock, while Zen series was faster in every way, and competitive with monolithic Raptorlake.

You say fancy things, you dress it up fancy, but it ends up being plane jane. How many people outside of Intel fans will see it otherwise?

They used to lead in DDR implementation, memory controller performance, and caches. Ah, but it's just the fault of the "process" right? RIGHT?

You're missing the point. RWC is weak. Foveros isn't.

 

[igor_kavinski]

Which was my problem with Igor's claim of Intel being #2.

They have better RT cores, do they not? Regardless of whether they are using more die area for that. I like the fact that if I suddenly developed an RT fetish, I could get an A770 or the Battlemage A870 and satisfy that urge without making Jensen richer.

 

[coercitiv]

Foveros is top tier packaging, it just happens to hang around with some very shady core types. I fear some day we'll find it face down in some organic substrate, a victim of circumstance. Such awesome potential, such a sad story.

 

[igor_kavinski]

It's equal to if Zen 5/Lion Cove reduces CPU performance by 20% but improves AI performance by 20x. Yay?

AMD would love that for bragging rights. Some previously infeasible AI applications would become a reality. So yeah, that would be a YAY!

 

[FlameTail]

Foveros is top tier packaging, it just happens to hang around with some very shady core types. I fear some day we'll find it face down in some organic substrate, a victim of circumstance. Such awesome potential, such a sad story.

Foveros, Lion Cove and Alchemist walk into a bar...

 

[SiliconFly]

Foveros, Lion Cove and Alchemist walk into a bar...

They get hit by an arrow and fall into a lake.

 

[Hulk]

Ok, but Meteorlake sometimes underperforms the predecessor at the same clock, while Zen series was faster in every way, and competitive with monolithic Raptorlake.

You say fancy things, you dress it up fancy, but it ends up being plane jane. How many people outside of Intel fans will see it otherwise?

They used to lead in DDR implementation, memory controller performance, and caches. Ah, but it's just the fault of the "process" right? RIGHT?

You have some good points here. I'll try to answer.

If Foveros is so great then why is there such a performance hit when moving from monolithic Raptor Cove to supposedly superior Redwood Cove and Foveros?

There will always be a performance loss/increase in latency when moving from monolithic to tiles. Intel slightly improved Redwood Cove to try and offset that in moving to Meteor Lake.

AMD made the move to tiles many generations ago so when moving through current generations they were on equal footing (all tiled).

But your original point is a good one. Tiled Zen 4 still had at least as good, probably better IPC than tiled Redwood Cove despite all this hype regarding Foveros. Why?

Couple of reasons. Zen 4 core is really good, better than we even thought as it "makes up for being tiled." Foveros isn't really doing anything special compared to AMD tiled solution.

AMD has had a couple of generation to get their tiled "act" together. This of course doesn't help Intel because you are right, they did a lot of talking and Meteor Lake is just "okay" in my opinion.

 

[jpiniero]

You have some good points here. I'll try to answer.

If Foveros is so great then why is there such a performance hit when moving from monolithic Raptor Cove to supposedly superior Redwood Cove and Foveros?

Meteor Lake is better than Raptor Lake. Just not all that much. That more likely has to do with Intel 4 than Foveros.

 

[SiliconFly]

Meteor Lake is better than Raptor Lake. Just not all that much. That more likely has to do with Intel 4 than Foveros.

TBH, MTL isn't that much better than RPL in performance. It should have been more. A lot more.

And with RWC, Intel was focused more on improving it's efficiency rather than performance. And both didn't pan out well if you ask me. Like one forum member aptly said about RWC: "they just put lipstick on a p*g and called it a day". Sadly, RWC is just RPC/GLC with some minor efficiency tweaks. Nothing more.

 

[FlameTail]

You have some good points here. I'll try to answer.

If Foveros is so great then why is there such a performance hit when moving from monolithic Raptor Cove to supposedly superior Redwood Cove and Foveros?

There will always be a performance loss/increase in latency when moving from monolithic to tiles. Intel slightly improved Redwood Cove to try and offset that in moving to Meteor Lake.

AMD made the move to tiles many generations ago so when moving through current generations they were on equal footing (all tiled).

But your original point is a good one. Tiled Zen 4 still had at least as good, probably better IPC than tiled Redwood Cove despite all this hype regarding Foveros. Why

Couple of reasons. Zen 4 core is really good, better than we even thought as it "makes up for being tiled." Foveros isn't really doing anything special compared to AMD tiled solution.

AMD has had a couple of generation to get their tiled "act" together. This of course doesn't help Intel because you are right, they did a lot of talking and Meteor Lake is just "okay" in my opinion.

Why are you calling AMD chiplets as 'tiles'?

 

[Hulk]

Why are you calling AMD chiplets as 'tiles'?

View attachment 93944

You're right I used the wrong term but the point is both AMD and Meteor Lake are disaggregated, that's what I meant.

 

[Hulk]

Meteor Lake is better than Raptor Lake. Just not all that much. That more likely has to do with Intel 4 than Foveros.

Better in what ways?

Efficiency, then yes of course.

IPC? I don't think so but I could be wrong? I was under the impression that Raptor Lake does better than Meteor at iso frequency?

 

[Hitman928]

Better in what ways?

Efficiency, then yes of course.

IPC? I don't think so but I could be wrong? I was under the impression that Raptor Lake does better than Meteor at iso frequency?

IPC I think is basically equal across a large testing suite. Loads that are really latency sensitive probably perform worse on MTL clock for clock.

 

[trivik12]

https://twitter.com/x/status/1758437768185979123

He has been leaking some intel related stuff recently. Another guy from India. huge growth in NPU Tops and he is also saying overall will have 100 TOPS worth power. That is nuts for sure.

 

[adroc_thurston]

Latest advanced 3D packaging technologies like foveros

It's not 3D, the base is passive.

and outdated substrate based 2.5D tech like in Zen5.

That's 2D flipchip.
2.5D is passive slabs a-la MTL.

MTL has a brand new NOC.

It's IOSF. From 2012.
You can even like look for MTL hotchips slideware.

If Foveros is so great then why is there such a performance hit when moving from monolithic Raptor Cove to supposedly superior Redwood Cove and Foveros?

Because they moved the relevant bits (IMC and pretty much the rest of the uncore) from the ring to IOSF.