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

Page 293 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Tigerick

Senior member
Apr 1, 2022
683
563
106
PPT1.jpg
PPT2.jpg
PPT3.jpg



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

ModelCode-NameDateTDPNodeTilesMain TileCPULP E-CoreLLCGPUXe-cores
Core Ultra 100UMeteor LakeQ4 202315 - 57 WIntel 4 + N5 + N64tCPU2P + 8E212 MBIntel Graphics4
?Lunar LakeQ4 202417 - 30 WN3B + N62CPU + GPU & IMC4P + 4E08 MBArc8
?Panther LakeQ1 2026 ??Intel 18A + N3E3CPU + MC4P + 8E4?Arc12



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

Meteor LakeArrow Lake (20A)Arrow Lake (N3B)Arrow Lake Refresh (N3B)Lunar LakePanther Lake
PlatformMobile H/U OnlyDesktop OnlyDesktop & Mobile H&HXDesktop OnlyMobile U OnlyMobile H
Process NodeIntel 4Intel 20ATSMC N3BTSMC N3BTSMC N3BIntel 18A
DateQ4 2023Q1 2025 ?Desktop-Q4-2024
H&HX-Q1-2025
Q4 2025 ?Q4 2024Q1 2026 ?
Full Die6P + 8P6P + 8E ?8P + 16E8P + 32E4P + 4E4P + 8E
LLC24 MB24 MB ?36 MB ??8 MB?
tCPU66.48
tGPU44.45
SoC96.77
IOE44.45
Total252.15



Intel Core Ultra 100 - Meteor Lake

INTEL-CORE-100-ULTRA-METEOR-LAKE-OFFCIAL-SLIDE-2.jpg

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.)

Clockspeed.png
 

Attachments

  • PantherLake.png
    PantherLake.png
    283.5 KB · Views: 23,970
  • LNL.png
    LNL.png
    881.8 KB · Views: 25,442
Last edited:

S'renne

Member
Oct 30, 2022
136
99
61
In spite of everything, MTL performance is still only on par with previous gen (and that was a bit of a disappointment for me).

And it's battery life is also only on par with the competition (in spite of all the power saving tech). Well, there are a few scenarios where it does provide better results. But there is no clear win and that speaks volumes.



XMX is pretty much necessary these days but isn't mandatory. MTL can definitely live without it.




I don't think it's a fair comparison. When AMD released it's NPU, there weren't many mainstream AI apps to begin with. Today, we have enough apps and more mainstream apps are integrating AI/LLMs as we speak. Things are just getting started and AMD was actually on the leading edge. Forget day 0 for a moment, today all those CPUs support AI and that matters a lot.
What...AMD's Xillinx NPU still does nothing unlike Intel now its not even a day 0 problem. Besides that ROCm kind of works only on Linux either, and they're already known for firmware/software support issues on their GPUs too for AI/ML whereas Intel already have a suite for Stable Diffusion, they also did a demo on AI assisting audio workstation tasks with a Lunar Lake sample chip.

Hardware is leading edge yes, software support is the worst
 

deasd

Senior member
Dec 31, 2013
526
800
136
After the patch which enable the 'stable mode' on RPL i9 and led to performance decline,




I wonder if this old ARL performance projection still make sense? ARL is 20%faster than RPL-R, and how much does ARL faster than RPL now if RPL have to be under 'stable mode'?


Image-01-980x346.jpg
 

S'renne

Member
Oct 30, 2022
136
99
61
After the patch which enable the 'stable mode' on RPL i9 and led to performance decline,




I wonder if this old ARL performance projection still make sense? ARL is 20%faster than RPL-R, and how much does ARL faster than RPL now if RPL have to be under 'stable mode'?


View attachment 97349
No idea actually since we dont even know if there's variance between 20A and 3NB nodes for Arrow Lake nor the memory configurations for the benchmarks among other metrics
 

inf64

Diamond Member
Mar 11, 2011
3,708
4,056
136
After the patch which enable the 'stable mode' on RPL i9 and led to performance decline,




I wonder if this old ARL performance projection still make sense? ARL is 20%faster than RPL-R, and how much does ARL faster than RPL now if RPL have to be under 'stable mode'?


View attachment 97349
20% in *some* MT workloads ,and 5-10% tops in ST workloads vs the old RPL-R (another slide claimed ~5% ST uplift vs RPL). My guess is it will get crushed by Granite Ridge, but let's wait and see.
 

mikk

Diamond Member
May 15, 2012
4,155
2,165
136
What's the score of a 253W 14900K, is it more than 35.8K? You have to assume reviewers will be using the new baseline profile as default and it's Asus only afaik.
 

S'renne

Member
Oct 30, 2022
136
99
61
20% in *some* MT workloads ,and 5-10% tops in ST workloads vs the old RPL-R (another slide claimed ~5% ST uplift vs RPL). My guess is it will get crushed by Granite Ridge, but let's wait and see.
Im just hoping its not getting crushed- well not entirely according to All The Watts leaks
 

Ghostsonplanets

Senior member
Mar 1, 2024
410
694
96
Another MTL-U review in the wild. Once again by NotebookCheck:


TLDR: Lackluster CPU that can be slower than last year model. Unimpressive upgrade all around. 4 Xe/64 EU tGPU fails to surpass last year Iris Xe 96EU, being, on average, on par with last year GFX performance. Battery life tests are incomplete, but Notebook Check reports nearly identical 10 hour battery life on WLAN surfing test (With one screen disabled) as last year RPL-U model.

Notebook Check still recommended the machine because there isn’t much difference in price between last year and current model. But if the last year model drops in price, it's recommended to buy the last year one as MTL-U fails to raise the baseline over RPL-U all around.
 

Ghostsonplanets

Senior member
Mar 1, 2024
410
694
96
Also, I've been looking at Core Ultra 5 115U page on Intel Ark and it still lacks a suggested client pricing, despite being claimed that it was "Launched" and introduced in Q4'23🤔.

Either we'll see it on cheaper back to school laptops in H2 or Intel doesn't plan any volume at all for it.
 
Last edited:
  • Like
Reactions: spursindonesia

mikk

Diamond Member
May 15, 2012
4,155
2,165
136
Another MTL-U review in the wild. Once again by NotebookCheck:


TLDR: Lackluster CPU that can be slower than last year model. Unimpressive upgrade all around. 4 Xe/64 EU tGPU fails to surpass last year Iris Xe 96EU, being, on average, on par with last year GFX performance. Battery life tests are incomplete, but Notebook Check reports nearly identical 10 hour battery life on WLAN surfing test (With one screen disabled) as last year RPL-U model.

Same Witcher 3 performance at 30W versus 45W on the previous model with Raptor Lake. Notebookcheck should use something with a more realistic load for the CPU power test like Cinebench where we have also have benchmarks results to compare. I mean who plays with Prime95, we don't even have performance numbers from this.
 

Abwx

Lifer
Apr 2, 2011
11,067
3,732
136
Same Witcher 3 performance at 30W versus 45W on the previous model with Raptor Lake. Notebookcheck should use something with a more realistic load for the CPU power test like Cinebench where we have also have benchmarks results to compare. I mean who plays with Prime95, we don't even have performance numbers from this.

Both laptops run at 40W sustained in Witcher 3 measured at the main, beside the 1355U has a 16W base TDP although it can sustain 52W for a short duration in this game (measured at the main).

NBC use difficultly comparable laptops settings wise but on the long run a 6 + 8 MTL U should be somewhat better than the previous 6 + 4 U gen even if it didnt live to expectations, here it use about 8-9W less at the SoC level compared to RPL, so that s something like 24W vs 32-33W SoC power.
 
  • Like
Reactions: spursindonesia

Hulk

Diamond Member
Oct 9, 1999
4,274
2,097
136
Let's use the Geekbench ST and MT performance below to "back into" ARL 15900K or whatever it will be called.
Assume 8+16 part and no HT.

I am also assuming a 5% clock regression with the move from 13900K to 15900K, or whatever it will be called.
Lion Cove 5.4GHz ST, 5.2GHz MT
Skymont would be 4.1GHz

After clockspeed regression let's assume Lion Cove will be 13% more performant than RPC (taking the higher ST range).
After clockspeed regression let's assume the MT score will be 15.5% better than Raptor Lake, lower end of the range indicated.

Drum roll please!

Lion Cove will need an IPC increase of 19% over Raptor Cove. Okay, seems reasonable. 5.4GHz ST and 5.2GHz all-core with a 19% IPC increase over Raptor Lake. It's a tall order when you factor in the tile latency but definitely in the realm of possibility.

Now let's consider Skymont. This is where we run into trouble.

Skymont at 4.1GHz will need an IPC increase of 60% over Gracemont in order to achieve the final MT score. The huge boost required for Skymont is due to the lack of HT in Lion Cove.

Now Skymont may run at 4.3GHz and that would help a bit, but to pull off a 15% MT increase over Raptor Lake with a 5% clock regression, Lion Cove will need 19% IPC increase and even at 4.3GHz Skymont will need a 50% minimum IPC increase over Gracemont. Seems unrealistic to me. I could see 30% or even 35% but I don't see 50%

This is why I've been saying the way out of this MT problem for Intel is to simply increase ARL to 8+24. If the transistor density is there to make this economically feasible.

 

mikk

Diamond Member
May 15, 2012
4,155
2,165
136
Both laptops run at 40W sustained in Witcher 3 measured at the main, beside the 1355U has a 16W base TDP although it can sustain 52W for a short duration in this game (measured at the main).

This is not what the test says.

Running Witcher 3 draws about 30 W compared to 45 W on the 2023 model. The newer Arc 4 GPU can therefore provide the same graphics performance as the Iris Xe 96 EUs but at significantly lower power requirements for superior performance-per-watt. This likely explains the quieter-than-expected fan noise levels mentioned above.
 

Rigg

Senior member
May 6, 2020
472
979
136
What's the score of a 253W 14900K, is it more than 35.8K? You have to assume reviewers will be using the new baseline profile as default and it's Asus only afaik.

TPU covered this pretty well. The initial review has scores with power limits removed, a 5.5 OC, and stock. They also did a follow up article with various power limits.


cinebench-multi.png


cinebench-multi.png

 
  • Like
Reactions: mikk

Abwx

Lifer
Apr 2, 2011
11,067
3,732
136
This is not what the test says.

They have the power comsumption curves at the main displayed, in Witcher 3 the 1355U laptop consume 52-53W at the start while MTL U is at about 40W, 12-13W difference at the main can in no way amount to 15W difference at the SoC level, more like 9W, so if MTL U is at say its 24W base TDP it means that the 1355 U is at 33W, and that s assuming that the laptop use 40W for 24W SoC power.
 

Ghostsonplanets

Senior member
Mar 1, 2024
410
694
96
Is it known how does Meteor Lake U treats their 2P cores with regards to clock and scheduling?

Meteor Lake H has 2P cores that use 8VT cells and are treated as Super Cores, similarly to QCOM Prime core idea. They can clock higher and are parked once you're using battery power. With the others 4P cores using 6VT cells and being usable either on battery power or charger power.

How would do that work with MTL-U 2P cores? 1P core using 8VT cells and being parked while on battery power? Both using 6VT cells due to power efficiency targets?

Would make sense if the Super P Core idea is something exclusive to MTL-H due to having many P cores. MTL-U only has 2P cores to work. Certainly would make for a more simplified scheduling logic.
 

Hitman928

Diamond Member
Apr 15, 2012
5,407
8,309
136
Is it known how does Meteor Lake U treats their 2P cores with regards to clock and scheduling?

Meteor Lake H has 2P cores that use 8VT cells and are treated as Super Cores, similarly to QCOM Prime core idea. They can clock higher and are parked once you're using battery power. With the others 4P cores using 6VT cells and being usable either on battery power or charger power.

How would do that work with MTL-U 2P cores? 1P core using 8VT cells and being parked while on battery power? Both using 6VT cells due to power efficiency targets?

Would make sense if the Super P Core idea is something exclusive to MTL-H due to having many P cores. MTL-U only has 2P cores to work. Certainly would make for a more simplified scheduling logic.

I’ve seen this idea of MTL having P cores and super P cores a few times now but I’ve never seen any sort of documentation or confirmation from Intel that this is true. Do you have a source for this?
 

Ghostsonplanets

Senior member
Mar 1, 2024
410
694
96
I’ve seen this idea of MTL having P cores and super P cores a few times now but I’ve never seen any sort of documentation or confirmation from Intel that this is true. Do you have a source for this?

I haven't seen documentation from Intel stating this either. But it was tested by independent reviewers (Chinese netizens + some knowledgeable people here in the west) and MTL does have a differentiation between Super P Core and normal P cores in scheduling. Some of the testing also found out that Super P Cores are parked while on battery, just like LPE cores are while on charger power.



GCu2dCGbwAAkMH9.jpeg



MTL has two types of P-core: 8VT/6VT, one with a high upper frequency limit and the other with low leakage

From "ThunderShock":

MTL has a different turbo design.
In the old product, Intel selects two good cores and turbo it.
But in the MTL, it has two types P-cores, ULT and XLT, not the same achieved. ULT can get more freq in equal voltage in high freq and only 2 cores, but not good in low freq.

Geekerwan:
Screenshot_20240421_125207_YouTube.jpg

"Super large core performance library* x 2"

"Large core density library* x 4"

* Performance and Density library refer to 8VT and 6VT cells as Intel 4 doesn't feature HD/UHD libraries as they're commonly referred.
 

Hitman928

Diamond Member
Apr 15, 2012
5,407
8,309
136
I haven't seen documentation from Intel stating this either. But it was tested by independent reviewers (Chinese netizens + some knowledgeable people here in the west) and MTL does have a differentiation between Super P Core and normal P cores in scheduling. Some of the testing also found out that Super P Cores are parked while on battery, just like LPE cores are while on charger power.



View attachment 97414





From "ThunderShock":



Geekerwan:
View attachment 97413

"Super large core performance library* x 2"

"Large core density library* x 4"

* Performance and Density library refer to 8VT and 6VT cells as Intel 4 doesn't feature HD/UHD libraries as they're commonly referred.

Seems like a bunch of guess work with very shaky evidence. Seems highly unlikely to me that Intel would take the time and resources to design and validate a 4th core type for MTL for what, a few percent performance increase for single threaded loads and increased power consumption in all loads? Their perf/w chart also doesn't correlate with the chart Intel showed when comparing their 6VT and 8VT cells for power/freq. Lastly, if they are using a higher density library for the "regular P-cores" and a lower density library for the "super P-cores", this should be observable in the die shot analyses that have been performed, but from everything I've seen, all of the P-cores are the same size and look identical.

I'd need to see much more convincing evidence to believe that there are actually different P-cores on the MTL die rather than it being normal wafer gradient causing characteristic differences between the cores on the same die. This might be more of a problem on Intel 4 than on previous nodes, leading to their speculation.
 

Khato

Golden Member
Jul 15, 2001
1,206
251
136
They have the power comsumption curves at the main displayed, in Witcher 3 the 1355U laptop consume 52-53W at the start while MTL U is at about 40W, 12-13W difference at the main can in no way amount to 15W difference at the SoC level, more like 9W, so if MTL U is at say its 24W base TDP it means that the 1355 U is at 33W, and that s assuming that the laptop use 40W for 24W SoC power.
They do? I only see the power consumption curves for the 1355U variant. Along with an Editor's Note stating that the section will be updated once the multimeter used for collecting the data is repaired.
 

Abwx

Lifer
Apr 2, 2011
11,067
3,732
136

SiliconFly

Golden Member
Mar 10, 2023
1,069
554
96
I’ve seen this idea of MTL having P cores and super P cores a few times now but I’ve never seen any sort of documentation or confirmation from Intel that this is true. Do you have a source for this?
Still rumors. Not yet confirmed.

Seems like a bunch of guess work with very shaky evidence. Seems highly unlikely to me that Intel would take the time and resources to design and validate a 4th core type for MTL for what, a few percent performance increase for single threaded loads and increased power consumption in all loads?
Totally agree. Having 2 different P core designs is a complete waste of time and money.
 

mikk

Diamond Member
May 15, 2012
4,155
2,165
136
Let's use the Geekbench ST and MT performance below to "back into" ARL 15900K or whatever it will be called.
Assume 8+16 part and no HT.

I am also assuming a 5% clock regression with the move from 13900K to 15900K, or whatever it will be called.
Lion Cove 5.4GHz ST, 5.2GHz MT
Skymont would be 4.1GHz

After clockspeed regression let's assume Lion Cove will be 13% more performant than RPC (taking the higher ST range).
After clockspeed regression let's assume the MT score will be 15.5% better than Raptor Lake, lower end of the range indicated.

Drum roll please!

Lion Cove will need an IPC increase of 19% over Raptor Cove. Okay, seems reasonable. 5.4GHz ST and 5.2GHz all-core with a 19% IPC increase over Raptor Lake. It's a tall order when you factor in the tile latency but definitely in the realm of possibility.

Now let's consider Skymont. This is where we run into trouble.

Skymont at 4.1GHz will need an IPC increase of 60% over Gracemont in order to achieve the final MT score. The huge boost required for Skymont is due to the lack of HT in Lion Cove.

Now Skymont may run at 4.3GHz and that would help a bit, but to pull off a 15% MT increase over Raptor Lake with a 5% clock regression, Lion Cove will need 19% IPC increase and even at 4.3GHz Skymont will need a 50% minimum IPC increase over Gracemont. Seems unrealistic to me. I could see 30% or even 35% but I don't see 50%

This is why I've been saying the way out of this MT problem for Intel is to simply increase ARL to 8+24. If the transistor density is there to make this economically feasible.



From the Lunar Lake die picture it clearly looks like Skymont grow exponentially compared to Lion Cove. The 4 core Skymont cluster including L2 is closer to 2 Lion Cove cores in this. With Gracemont it was closer to 1 Golden Cove core.
 

SiliconFly

Golden Member
Mar 10, 2023
1,069
554
96
From the Lunar Lake die picture it clearly looks like Skymont grow exponentially compared to Lion Cove. The 4 core Skymont cluster including L2 is closer to 2 Lion Cove cores in this. With Gracemont it was closer to 1 Golden Cove core.
Or is it that the E cores are still similar in size and the new P cores are smaller instead? Just thinking...

Is there any link to that lunar lake die shot?

(Note: I still haven't seen lunar lake die shot yet.)
 

mikk

Diamond Member
May 15, 2012
4,155
2,165
136
Or is it that the E cores are still similar in size and the new P cores are smaller instead? Just thinking...

Is there any link to that lunar lake die shot?

(Note: I still haven't seen lunar lake die shot yet.)

From here



Lion Cove appears slightly smaller than Redwood Cove but this is a comparison with Intel 4 to TSMC N3B. Skymont size on the other side increased ~50% with a more dense node.