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.)

 

[TESKATLIPOKA]

And no Armoury Crate. Lenovo Vantage is optional, whereas AC finds a way into your system.

I find Lenovo Vantage pretty decent and didn't uninstall It. If I need something I start It then close It.
Just be aware that USB-C ports can be misaligned, at least I have this problem, but they are still working fine, just a visual defect.

 

[cebri1]

We still need more test to confirm new pcode situation (if confirmed new BIOS includes the update) but also the new GPU drivers which should fix the abysmal drop in performance of the iGPU when going under 30W. If Intel manages to fix those things, MTL will become a very interesting product for 24.

 

[TESKATLIPOKA]

If Intel manages to fix those things, MTL will become a very interesting product for 24.

A decent product, but for 2024 the interesting products should be the next gen -> Strix Point and ARL.
Although not so sure what to expect from ARL, because from leaks It doesn't look like a big IPC increase can be expected and who knows about core count. At least with Strix we have 50% more cores and >>10% IPC gain.

 

[cebri1]

A decent product, but for 2024 the interesting products should be the next gen -> Strix Point and ARL.
Although not so sure what to expect from ARL, because from leaks It doesn't look like a big IPC increase can be expected and who knows about core count. At least with Strix we have 50% more cores and >>10% IPC gain.

Those are going to be 2H products for sure. Probably Q4. Lunar Lake is in my opinion more interesting than ARL-H.

 

[TESKATLIPOKA]

Those are going to be 2H products for sure. Probably Q4. Lunar Lake is in my opinion more interesting than ARL-H.

Yeah, we will have to wait for them to be released, but performance uplift will be significant.
Lunar Lake will be interesting for It's efficiency, new IGP and maybe 1T, but I am not sure how good It will perform in nT thanks to only having 4P+4E+2LPE.

 

[cebri1]

Yeah, we will have to wait for them to be released, but performance uplift will be significant.
Lunar Lake will be interesting for It's efficiency, new IGP and maybe 1T, but I am not sure how good It will perform in nT thanks to only having 4P+4E+2LPE.

I would add on-die DDR5, which should improve iGPU (and inference) performance a lot.

 

[TESKATLIPOKA]

I would add on-die DDR5, which should improve iGPU (and inference) performance a lot.

How would that improve It?
It's just 16-32GB 8533mbps LPDDR5 on package. It won't perform better than If It was on board. Any performance gain will be because It's faster.

 

[cebri1]

How would that improve It?
It's just 16-32GB 8533mbps LPDDR5 on package. It won't perform better than If It was on board. Any performance gain will be because It's faster.

I would expect a wider bus and an increase in effective bandwidth. If not... why move the RAM to the chip?

At least that is what I interpreted from the leaks

 

[TESKATLIPOKA]

I would expect a wider bus and an increase in effective bandwidth. If not... why move the RAM to the chip?

With wider bus you can't have 16 or 32GB unless It's 256-bit instead of 128-bit. That 128EU IGP doesn't need that much BW, It would be enough even for 256-320EU IGP.
The reason why It is moved there is because It saves space on the motherboard in my opinion and probably a bit of power.

 

[Geddagod]

Yeah, we will have to wait for them to be released, but performance uplift will be significant.
Lunar Lake will be interesting for It's efficiency, new IGP and maybe 1T, but I am not sure how good It will perform in nT thanks to only having 4P+4E+2LPE.

no LPE cores, just 4+4 it looks like

 

[FlameTail]

. At least with Strix we have 50% more cores

But a mix of Zen5 + Zen5C.

So the core count increase alone will not necessarily bring a 50% MT uplift.

 

[FlameTail]

I would add on-die DDR5, which should improve iGPU (and inference) performance a lot.

IT IS NOT ON-DIE.

IT IS ON-PACKAGE.

 

[FlameTail]

With wider bus you can't have 16 or 32GB unless It's 256-bit instead of 128-bit. That 128EU IGP doesn't need that much BW, It would be enough even for 256-320EU IGP.
The reason why It is moved there is because It saves space on the motherboard in my opinion and probably a bit of power.

Pros of on-package LPDDR:

1. Space savings : You don't waste a ton of space on SODIMM/CAMM2 slots.

2. Power savings : Shorter traces can reduce power consumption (slightly).

3. Cost savings : Shorter traces and not having DIMM slots gives cost savings to laptop OEMs.

4. Performance improvement: Having the RAM closer to the processor reduces the trace length and in theory should reduce latency to RAM.

Cons of on-package RAM

1. Non-upgradeable.

 

[coercitiv]

But a mix of Zen5 + Zen5C.

So the core count increase alone will not necessarily bring a 50% MT uplift.

You should consider the different approach for making the dense cores.

Spoiler: Zen 4 vs. Zen 4c

 

[cebri1]

https://x.com/9550pro/status/1737779455077347329?s=20

MTL with dummy die

 

[mikk]

NBC measured power at the wall with the internal screen disabled, the laptop hoover at 83-85W at the start of Cinebench, you think that a laptop that pull this amount of power has a CPU that run at a fixed 28W..?..

I gave the wrong link, here the correct one: https://www.notebookcheck.net/Intel...-convinces-with-GPU-performance.783320.0.html

They set a fixed TDP to 28W, 35W, 45W in a special Cinebench R23 scaling test, your quotes are not related. They can't set fixed power limits on the wall and they refer to the TDP, it is not a power wall measurement.

 

[mikk]

A decent product, but for 2024 the interesting products should be the next gen -> Strix Point and ARL.
Although not so sure what to expect from ARL, because from leaks It doesn't look like a big IPC increase can be expected and who knows about core count. At least with Strix we have 50% more cores and >>10% IPC gain.

ARL-H and Strix Point will be 2025 contenders and Meteor Lake is a 2024 contender. Yes even if they are launching it at the end of 2024 or let's say Q4 it's similar to MTL. On AMD it usually takes a much longer time bringing them out into the market after the launch.

As for ARL-H the worrying thing is that hey are reusing the Soc and tile layout with all its flaws. Also the tGPU will be largely reused (slightly updated Alchemist). Lunar Lake is a real next gen Soc but this is limited to 4+4.

 

[Abwx]

I gave the wrong link, here the correct one: https://www.notebookcheck.net/Intel...-convinces-with-GPU-performance.783320.0.html

They set a fixed TDP to 28W, 35W, 45W in a special Cinebench R23 scaling test, your quotes are not related. They can't set fixed power limits on the wall and they refer to the TDP, it is not a power wall measurement.

I saw this curve, they made a mistake because there s no power/frequency curve that can be convexe like this one, all are inherently concave, that is, perf/watt decrease as power increase while here it s the contrary between 35 and 45W.

It s biaised because the turbo is working and shift the measurement.

From 35W to 45W the score increase by 20.6%, that s impossible as 20.6% higher frequency require something like 50% more power, the theorical best case with perfect mosfets that do no exist would be (1.206)^2 = 1.45x more power for such a frequency uplift.

At 45W it should score about 13500 pts assuming the number at 28W is accurate.

 

[mikk]

I saw this curve, they made a mistake because there s no power/frequency curve that can be convexe like this one, all are inherently concave, that is, perf/watt decrease as power increase while here it s the contrary between 35 and 45W.

It s biaised because the turbo is working and shift the measurement.

From 35W to 45W the score increase by 20.6%, that s impossible as 20.6% higher frequency require something like 50% more power, the theorical best case with perfect mosfets that do no exist would be (1.206)^2 = 1.45x more power for such a frequency uplift.

At 45W it should score about 13500 pts asuming the number at 28W is accurate.

Scores are too low, there is no turbo involved. By setting a fixed power number there is no PL2 involved, this the purpose of such a scaling test.

MTL seems to scale quite good with the power yes, even though it might got a little screwed because of the pcode issue. It is even worse in GPU benchmarks below 25W which is partially a driver issue says golden pig. I would not draw a conclusion at the moment, it definitely needs the pcode update and driver update for GPU benchmarks as well.

 

[Abwx]

Scores are too low, there is no turbo involved. By setting a fixed power number there is no PL2 involved, this the purpose of such a scaling test.

MTL seems to scale quite good with the power yes, even though it might got a little screwed because of the pcode issue. It is even worse in GPU benchmarks below 25W which is partially a driver issue says golden pig. I would not draw a conclusion at the moment, it definitely needs the pcode update and driver update for GPU benchmarks as well.

Whatever the way it scale it power cant scale quasi linearly from 35W to 45W like in this weird curve, power increase at least as a square of frequency, moreover in this range where the uncore power is no more significant, i m talking of CPU perf for the time.

So if the power/frequency curve is convexe or even linear it means that data acquisition was flawed, we ll undoubtly have confirmation once commercial and final models are released.

 

[TESKATLIPOKA]

But a mix of Zen5 + Zen5C.

So the core count increase alone will not necessarily bring a 50% MT uplift.

It's not a problem that It's a mix, because the core is the same, only V/F curve is different, but It's actually advantageous for limited TDP as @coercitiv posted.
The only question is how big an uplift you will get at lower power limits.

Pros of on-package LPDDR:

1. Space savings : You don't waste a ton of space on SODIMM slots.

2. Power savings : Shorter traces can reduce power consumption (slightly).

3. Cost savings : Shorter traces and not having DIMM slots gives cost savings to laptop OEMs.

4. Performance improvement: Having the RAM closer to the processor reduces the trace length and in theory should reduce latency to RAM.

Cons of on-package RAM

1. Non-upgradeable.

LPDDR doesn't use DIMM slots, the memory is soldered on board.

 

[FlameTail]

LPDDR doesn't use DIMM slots, the memory is soldered on board.

Yeah, the first point was comparing with DDR. But then there is CAMM2 for LPDDR!

 

[TESKATLIPOKA]

no LPE cores, just 4+4 it looks like

You are right, I mistook It because somewhere on this forum was a debate about It, but I didn't pay much attention to It.
From diagram It looks like only P and E-cores.

 

[TESKATLIPOKA]

Yeah, the first point was comparing with DDR. But then there is CAMM2 for LPDDR!

I don't think CAMM2 will replace soldered memory. I think It will replace SO-DIMM.

 

[SiliconFly]

You are right, I mistook It because somewhere on this forum was a debate about It, but I didn't pay much attention to It.
From diagram It looks like only P and E-cores.

Aligns well with the twitter leak. Actually there are no E cores. It has 4 P cores & 4 LP E cores.