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.)

 

[cebri1]

And to be honest, I'm waiting for post-CES reviews but if Golden Pig pcore update results are validated it would beat Phoenix at perf/watt above 35W. Ofc is not the ideal scenario for a laptop CPU but it's much better than 13th gen against Phoenix/Zen4.

 

[Abwx]

You mean this?

Then Phoenix at 5W is comparable, at 6W is better. E-core at <5W is better.
I really want to know how he measured It.
I have PHX laptop and Ryzen Master is not working, the same for AMD Overdrive. Not sure about MTL and Extreme Tuning Utility.

That s made up numbers, he use the package power, so the uncore power is included in the alleged core power comsumption, that s why the curves are that steep at the beginning, by increasing artificially the idle power he shift the curves where it stand his doctored numbers.

Yet HVinfo has power per core data, one has to wonder why he didnt isolate a single core power using this tool, likely that it would had contradicted his purposely biaised numbers.

 

[DrMrLordX]

No it didn't. Even in the first Notebookcheck review, which everyone used to tell us MTL was a failure, it beat the 13th gen at all tested power levels.

There was SPECInt data on the same page where I posted showing Meteor Lake losing perf/watt versus a lowly 7840u from a year ago.

"It should clearly win ST performance versus its competition and older Intel SoCs, and it should use at least a little less power doing so"

And, not or. There are also some circumstances where Meteor Lake loses to Raptor Lake where it clearly should not. But if I need to clarify, then let me restate it in a way that everyone can understand:

Meteor Lake is supposed to beat Phoenix AND Raptor Lake. It clearly did not do so.

Now let's talk about how well Meteor Lake is doing on the desktop. Oh no, can't do that, sorry. Because it doesn't exist (for good reason).

One word: Linux.

Okay? Care to explain how Phoenix won out in the AI testbed? And no it wasn't Hawk Point either.

But Intel did not release a chip with the same performance as Zen 4 but half the power, did they?

Not even close!

 

[SiliconFly]

You mean this?

Then Phoenix at 5W is comparable, at 6W is better. E-core at <5W is better.
I really want to know how he measured It.
I have PHX laptop and Ryzen Master is not working, the same for AMD Overdrive. Not sure about MTL and Extreme Tuning Utility.

Looks interesting. Now waiting for CES for more MTL laptop reviews.

 

[tamz_msc]

Okay? Care to explain how Phoenix won out in the AI testbed? And no it wasn't Hawk Point either.

The author explains why: AVX-512

Also, does the NPU even have working drivers under Linux?

 

[DrMrLordX]

The author explains why: AVX-512

Kinda makes Intel look stupid after pushing the AI capabilities of Meteor Lake so hard.

Also, does the NPU even have working drivers under Linux?

Yes:

GitHub - intel/linux-npu-driver: Intel® NPU (Neural Processing Unit) Driver

Intel® NPU (Neural Processing Unit) Driver. Contribute to intel/linux-npu-driver development by creating an account on GitHub.

github.com

 

[Hitman928]

One word: Linux.

Alder Lake also had sub-par performance at launch on Linux.

MTL performs better in Linux than Windows.

 

[cebri1]

Meteor Lake is supposed to beat Phoenix

Not at all. MTL was supposed to close the gap as much as possible as noone in their right mind thought that Intel 4 + RWC (or GC+) was going to beat N4 + Zen4 in perf per watt. Although it might at some power levels.

Now let's talk about how well Meteor Lake is doing on the desktop. Oh no, can't do that, sorry. Because it doesn't exist (for good reason).

Why would you bring to desktop an architecture meant for power effiency, with exactly the same p-core as Raptor Lake and with a ST regression due to lower clocks?

MTL had three objetives:
- Improved perf/watt CPU over RPL. Target 20% from Intel4 + e-core IPC gains -> Looking good but pending more independent reviews.
- Provide a compelling iGPU offering -> As of today, best in the market. (you missed that phoronix benchmark for some reason).
- Integrate NPU in SOC -> Done

 

[coercitiv]

noone in their right mind thought that Intel 4 + RWC (or GC+) was going to beat N4 + Zen4 in perf per watt. Although it might at some power levels.

So some folks in this thread were crazy nuts at some power levels, yet perfectly sane at others?

 

[cebri1]

So some folks in this thread were crazy nuts at some power levels, yet perfectly sane at others?

"Meteor Lake is supposed to beat Phoenix". Implies MTL should/must beat Phoenix at all power levels -> crazy nuts
"oone in their right mind thought that Intel 4 + RWC (or GC+) was going to beat N4 + Zen4 in perf per watt. Although it might at some power levels." Implies Phoenix will more efficient specially at the W ranges more relevant for laptops and ultrathin books (15-35W), might beat it at power levels higher than 35-40W which is nice but defeats the purpose of MTL, power efficiency. -> perfectly sane.

 

[coercitiv]

"Meteor Lake is supposed to beat Phoenix". Implies MTL should/must beat Phoenix at all power levels -> crazy nuts
"oone in their right mind thought that Intel 4 + RWC (or GC+) was going to beat N4 + Zen4 in perf per watt. Although it might at some power levels." Implies Phoenix will more efficient specially at the W ranges more relevant for laptops and ultrathin books (15-35W), might beat it at power levels higher than 35-40W which is nice but defeats the purpose of MTL, power efficiency. -> perfectly sane.

Your rationale above also implies that the difference between crazy nuts and perfectly sane is 10-15W.

 

[cebri1]

Your rationale above also implies that the difference between crazy nuts and perfectly sane is 10-15W.

Nothing to do with that, it's a comparison between two power curves that do not touch ("MTL beats Phoenix") vs two power curves with a different slope that will eventually meet, before they met at 65W or so now it looks like they will meet at 40W. They are two complete different statements.

But anyway, anyone is free to say they were expecting Intel 4 to beat N4.

 

[SiliconFly]

Your rationale above also implies that the difference between crazy nuts and perfectly sane is 10-15W.

Those power curves don't include real world workloads that actually uses LPE cores. Real world efficiency and battery life of MTL i think is gonna be very different from what those "stress test curves" imply. Real world usage is vastly different. This CES gonna be very interesting for MTL I guess...

 

[H433x0n]

"oone in their right mind thought that Intel 4 + RWC (or GC+) was going to beat N4 + Zen4 in perf per watt. Although it might at some power levels."

Why? Intel 4 6VT might be a better node. I wouldn’t discount the effect of Zen 4’s cache efficiency (taking less cycles to receive, getting less misses) and the effect of a better branch predictor can’t be understated.

 

[Thunder 57]

I thought MTL launched almosta month ago now. How do we not have answers? Either because it didn't really launch or because it isn't all that great. Guess we'll know soon enough.

 

[DavidC1]

I thought MTL launched almosta month ago now. How do we not have answers? Either because it didn't really launch or because it isn't all that great. Guess we'll know soon enough.

Because it's laptop-only. Similar arguments back in the Icelake and Tigerlake era because they had no desktop apples-to-apples to compare.

 

[DrMrLordX]

Not at all. MTL was supposed to close the gap as much as possible as noone in their right mind thought that Intel 4 + RWC (or GC+) was going to beat N4 + Zen4 in perf per watt. Although it might at some power levels.

Who in their right mind sets performance targets like that?!?!? This isn't Cyrix or Zhaoxin VIA we're talking about. This is Intel! Meteor Lake was supposed to be a new uarch on a new node which gave Intel two opportunities to leapfrog the competition. Instead what we get is a mobile-only product that maybe sometimes beats the competition on perf/watt and can't outperform it decisively either.

Someone needs to be fired over that.

I thought MTL launched almosta month ago now. How do we not have answers? Either because it didn't really launch or because it isn't all that great. Guess we'll know soon enough.

I'm guessing the latter. Though given some of the delays on available models (3 weeks+) it may be both.

 

[ryanjagtap]

Just watched this video by Dr. Ian Cutress.

With respect to MTL showcasing Intel's IFS technologies, don't you find it weird that of the 5 die comprising MTL only the CPU tile is on Intel 4 with the base die on some old Intel node (which is ok as it just acts as an interposer), while the GPU tile is on TSMC 5nm, the I/O and SoC tile on TSMC 6nm. The Foveros tech used is also the older one with 45 micro bumps rather than the 36 micro bumps.

P.S: As the SoC tiles contain the LP-E cores, can we say that Intel uses TSMC for CPU in MTL? The LP-E cores are also given first priority in scheduling.

 

[eek2121]

ASUS may be announcing a NUC with MTL. We will have to wait and see.

Nothing to do with that, it's a comparison between two power curves that do not touch ("MTL beats Phoenix") vs two power curves with a different slope that will eventually meet, before they met at 65W or so now it looks like they will meet at 40W. They are two complete different statements.

But anyway, anyone is free to say they were expecting Intel 4 to beat N4.

The jury is still out on the process. I don’t think we will get our answer.

The new chip is designed in a new way that is causing odd performance regressions and other weirdness, so I wouldn’t use this to compare the process used.

Just watched this video by Dr. Ian Cutress.

With respect to MTL showcasing Intel's IFS technologies, don't you find it weird that of the 5 die comprising MTL only the CPU tile is on Intel 4 with the base die on some old Intel node (which is ok as it just acts as an interposer), while the GPU tile is on TSMC 5nm, the I/O and SoC tile on TSMC 6nm. The Foveros tech used is also the older one with 45 micro bumps rather than the 36 micro bumps.

P.S: As the SoC tiles contain the LP-E cores, can we say that Intel uses TSMC for CPU in MTL? The LP-E cores are also given first priority in scheduling.

Intel uses Samsung 14mm for their chipsets apparently (and have been for a while) if Raichu and others are correct.

Most of Intel’s EUV capacity is earmarked for 20a, 18a, and beyond. Because of this they have limited capacity.

 

[tamz_msc]

MTL performs better in Linux than Windows.

Yeah, because comparing apples to apples on Windows vs Linux with the same open-source compilers is a no go, because gcc on Windows is an afterthought.

 

[coercitiv]

P.S: As the SoC tiles contain the LP-E cores, can we say that Intel uses TSMC for CPU in MTL?

That would be like saying AMD is using ARM cores in Zen. The ARM core is also given priority in execution. Technically it's true, but it would be disingenuous to do so while ignoring the context.

 

[Tup3x]

Just watched this video by Dr. Ian Cutress.

With respect to MTL showcasing Intel's IFS technologies, don't you find it weird that of the 5 die comprising MTL only the CPU tile is on Intel 4 with the base die on some old Intel node (which is ok as it just acts as an interposer), while the GPU tile is on TSMC 5nm, the I/O and SoC tile on TSMC 6nm. The Foveros tech used is also the older one with 45 micro bumps rather than the 36 micro bumps.

P.S: As the SoC tiles contain the LP-E cores, can we say that Intel uses TSMC for CPU in MTL? The LP-E cores are also given first priority in scheduling.

One of the main point of that packing tech is that you can mix and match multiple different chips. If all would be made using same process... that doesn't make much sense.

 

[DrMrLordX]

One of the main point of that packing tech is that you can mix and match multiple different chips. If all would be made using same process... that doesn't make much sense.

It does if yields are too low on Intel 4 to produce a monolithic die of that size. Admittedly it makes more sense for a High NA EUV process with smaller reticle limits.

 

[SiliconFly]

...If all would be made using same process... that doesn't make much sense.

Actually, it doesn't matter even if the same process is used. The benefits are still the same. Smaller IPs managed by different/independent teams. Faster iteration & silicon validation. Hence faster to market.

It does if yields are too low on Intel 4 to produce a monolithic die of that size. Admittedly it makes more sense for a High NA EUV process with smaller reticle limits.

Also, making large monolithic dies are expensive due to lower yields due to their massive size. Smaller tiles reduce cost not just with higher yields, but also by shifting non-critical parts of the chip to cheaper and/or more suitable process nodes.

 

[SiliconFly]

Is ARL's tGPU based on an updated version MTL's Alchemist tGPU? Or is it based on Battlemage?

Recently, I read an article that said ARL's tGPU is Battlemage. I remember reading older articles that said it's based on Alchemist++. Any idea which one it's gonna be?