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.)

 

[coercitiv]

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

I think we're getting ahead of ourselves with semantics. If we follow Intel's marketing to the letter, there are no P cores in Lunar Lake either, there are power efficiency P cores. So it's 4 PE P cores & 4 LP E cores.

 

[FlameTail]

I think we're getting ahead of ourselves with semantics. If we follow Intel's marketing to the letter, there are no P cores in Lunar Lake either, there are power efficiency P cores. So it's 4 PE P cores & 4 LP E cores.

Ah screw this.

 

[FlameTail]

I think the future of laptop RAM is this:

(1) On-package LPDDR

or

(2) CAMM2 LPDDR

LPDDR soldered to the motherboard will die out. When the RAM is soldered, it's not upgradeable anyways, so why not instead go all the way and make it 'soldered on-package'. You get cost, space and performance benefits.

For laptops that need the upgradeable RAM, CAMM2 will be the way.

 

[mikk]

Arc on Linux runs faster than Windows 11.

When taking the geometric mean of all 84 tests that ended up running on each operating system, Ubuntu 23.10 with Linux 6.7 came out 15% faster than Windows 11.

Windows 11 vs. Ubuntu Linux Performance On The Intel Core Ultra 7 Meteor Lake Review - Phoronix

www.phoronix.com

Maybe related to the low power driver issue under Windows?

It's more a CPU test nevermind. Not much GPU tests.

 

[Hitman928]

Arc on Linux runs faster than Windows 11.

Windows 11 vs. Ubuntu Linux Performance On The Intel Core Ultra 7 Meteor Lake Review - Phoronix

www.phoronix.com

Maybe related to the low power driver issue under Windows?

It's more a CPU test nevermind. Not much GPU tests.

I think you mean MTL on Linux. . . most of those tests are CPU tests. In terms of graphics performance, the Linux side didn't quite match up with Windows 11.

Right now the integrated Intel Arc Graphics performance was around 85%+ the speed of the Windows 11 launch-day driver

 

[TESKATLIPOKA]

I think the future of laptop RAM is this:

(1) On-package LPDDR

or

(2) CAMM2 LPDDR

LPDDR soldered to the motherboard will die out. When the RAM is soldered, it's not upgradeable anyways, so why not instead go all the way and make it 'soldered on-package'. You get cost, space and performance benefits.

For laptops that need the upgradeable RAM, CAMM2 will be the way.

If memory is on package, then OEM is paying for It to Intel or AMD. Then OEM also loses better prices on that memory, because they don't have the same volume and the only thing OEM gains in return is some space saved.
If It was such a great thing, then It would have been used a long time ago.

 

[Glo.]

Arc on Linux runs faster than Windows 11.

Windows 11 vs. Ubuntu Linux Performance On The Intel Core Ultra 7 Meteor Lake Review - Phoronix

www.phoronix.com

Maybe related to the low power driver issue under Windows?

It's more a CPU test nevermind. Not much GPU tests.

Windows 11 vs. Ubuntu Linux Performance On The Intel Core Ultra 7 Meteor Lake Review - Phoronix

www.phoronix.com

Here is correct link for what you wanted to say. And no, its not faster, its actually slower than Windows performance.

 

[SiliconFly]

Phoronix claims MTL's iGPU wins in PPW against RDNA3 iGPU (780M)

https://twitter.com/x/status/1737491264516284491

Raja Koduri has retweeted this pretty much confirming the claim!

 

[igor_kavinski]

MTL should be even faster on Clear Linux.

 

[gdansk]

MTL should be even faster on Clear Linux.

Remains to be seen. Some of what Intel did on Clear Linux actually helps AVX-512 enabled CPUs more. In this case, their competitor's laptop CPU and not their own.

 

[igor_kavinski]

Remains to be seen.

I meant compared to Ubuntu so Windows 11 might seem even slower once Larabel re-runs the same benchmarks on Clear Linux.

 

[SpudLobby]

Idle and light load power is probably the strongest feature of MTL.

There are some issues currently with FW and OS scheduling which result in sub-optimal use of the LP E cores but this test does show the potential in idle and low use situations. This system here has a 120hz OLED BTW.

View attachment 90737

Utterly massive improvement over their previous products. Refresh every 15S I think isn’t as good as just a modest amount of cursor movement throughout the test demanding some constant input + new page every 30-45 seconds though. A big question for Intel is how low/high is the bar to keep the SoC tile exclusively running while retaining responsiveness.

To put it another way with a more demanding web browsing test with key and cursor input > 50% of the time I wonder if AMD starts to really win out. Here AMD does fine anyways even with Rembrandt and a just slightly smaller battery, but still.
It’s also impressive how well AMD is roughly doing with just a basic monolithic full 8c (vs Intel. They’re sloppy vs Apple).


Logically at some point what they end up back at taking that above point seriously is just building an SoC tile with like 4-6 low power, *good* performance cores from the ground up with a few “extra” cores for sheer MT and maybe ST on the compute tile, like even a 4 big core plus 4-8 general low power cores thing, because just two 2GHz cores blurs lines way too much and their compute tile is just such a disaster.

This is probably why Lunar Lake is like it is and drops all this crap, just has 4+4 on one tile.

 

[SpudLobby]

Power consumption as measured at the AC socket, yes. So including AC to DC power adapter, battery charge maintenance, etc. Simple fact is that without a power trace of CPU package power no definitive conclusion can be drawn regarding MTL efficiency versus entire laptop efficiency.

That’s ridiculous. AC to DC conversion included is real but not massive for this purpose. You can absolutely can draw a conclusion particularly over multiple systems, patterns and generalities will emerge. We don’t have that yet, but still.

Inb4 in a month or two: “This chip is actually uber efficient at idle, it just never shows up across 6 different OEMs and 23 different laptops, many with LPDDR5x” — and we all know Intel if anything does more OEM cooperation than AMD for the express purpose of ultrabooks, so.

I don’t know exactly what MTL idle looks like yet. One issue is we’re talking about a system with an entire gear switch from LP cores only to a compute tile + LP (SoC which is always on ofc), and we know their fabric for their main cores isn’t great for idle or low power.

So variance might be massive and finding where that line is for general use is what will be telling for better or worse. Entirely possible some results on mixed use are just a bit better than RPL, and others are right there with AMD’s Zen 4 mobile. Will depend on what the automated tests are like.

 

[PJVol]

This is probably why Lunar Lake is like it is and drops all this crap, just has 4+4 on one tile.

Indeed, and honestly I wonder how many more pages it will take for people to realize and agree that this CPU is simply pathetic with its "centrino moment"...
Idk, lol

 

[Khato]

That’s ridiculous. AC to DC conversion included is real but not massive for this purpose. You can absolutely can draw a conclusion particularly over multiple systems, patterns and generalities will emerge. We don’t have that yet, but still.

Is it? The notebookcheck idle power trace being referenced here with external monitor is claiming an average of 8.35W. The WiFi v1.3 battery life test result of 479 minutes on a 65Wh battery indicates an average power draw of 8.14W, and that's including the OLED screen at 150 cd/m^2 brightness. And the author specifically states that, " According to our energy measurements, it's the OLED panel that uses the most power when surfing." So it's pretty clear that the external monitor AC idle power trace is markedly higher than idle power on battery.

 

[Abwx]

So it's pretty clear that the external monitor AC idle power trace is markedly higher than idle power on battery.

That s because a lot of AC adaptators use a cheap power factor correction circuitry, basically a simple capacitor that is connected in parallel with the main, if the device power comsumption is low the reactive energy of the capacitor is drained by the main, wich increase the apparent comsumption.

Once the device use more than 10W or so the reactive energy is dumped in the adaptator circuitry and hence is used to supply the laptop.

So if apparent idle power is say 8W and that the laptop pull up to 85W it mean that the power delta is not 85 - 8 = 77W but more likely something like 80-81W.

 

[Khato]

That s because a lot of AC adaptators use a cheap power factor correction circuitry, basically a simple capacitor that is connected in parallel with the main, if the device power comsumption is low the reactive energy of the capacitor is drained by the main, wich increase the apparent comsumption.

Once the device use more than 10W or so the reactive energy is dumped in the adaptator circuitry and hence is used to supply the laptop.

So if apparent idle power is say 8W and that the laptop pull up to 85W it mean that the power delta is not 85 - 8 = 77W but more likely something like 80-81W.

For anyone who's interested in this subject I'd suggest a quick search on power factor correction of switched mode power supplies. We don't know whether notebookcheck is reporting active or apparent power, so power factor may or may not be at play. (Also of interest, residential metering is always active power.) As well, power factor doesn't change based on load.

 

[Abwx]

For anyone who's interested in this subject I'd suggest a quick search on power factor correction of switched mode power supplies. We don't know whether notebookcheck is reporting active or apparent power, so power factor may or may not be at play. (Also of interest, residential metering is always active power.) As well, power factor doesn't change based on load.

I once made some tests with a few 60-90W adaptators wich pulled 5-6W when not plugged on a device, once the capacitor was changed to a much lower value their idle power went to 2-4W depending of the brand.

As for NBC they dont report the numbers with such a precision because that s a negligible quantity in the respect of the tested devices average power comsumption.

 

[SpudLobby]

Is it? The notebookcheck idle power trace being referenced here with external monitor is claiming an average of 8.35W. The WiFi v1.3 battery life test result of 479 minutes on a 65Wh battery indicates an average power draw of 8.14W, and that's including the OLED screen at 150 cd/m^2 brightness. And the author specifically states that, " According to our energy measurements, it's the OLED panel that uses the most power when surfing." So it's pretty clear that the external monitor AC idle power trace is markedly higher than idle power on battery.

Okay yeah then it has to be about this. I wonder what’s going on then, and I wonder if a lot of the results under 20W that notebookcheck gets are a bit inflated.

 

[Hitman928]

For anyone who's interested in this subject I'd suggest a quick search on power factor correction of switched mode power supplies. We don't know whether notebookcheck is reporting active or apparent power, so power factor may or may not be at play. (Also of interest, residential metering is always active power.) As well, power factor doesn't change based on load.

If it’s reported in watts, then it’s active power.

 

[tamz_msc]

Running battery tests with an external monitor is not a good way to test because efficiency will depend on the internal circuitry of the connections that drive the external monitor (the port type) and the nature of the display engine of the SoC that drives the monitor.

For example, it will depend on refresh-rate how high/low the memory and other parts will clock when you connect the laptop to an external monitor.

Also, somebody like Apple, which makes products with use cases in mind that involve driving multiple displays, often very high resolution, over DP-on-USB-C/Thunderbolt - they leverage massive display engines that are "slow and wide", which nobody else in the industry does. This allows them to save more power versus the competition when set up identically with multiple displays.

The best way to measure power isolating the internal display is to set its brightness to zero.

All this is from a conversation I had with AT's erstwhile SSD reviewer Billy Tallis on reddit.

 

[SpudLobby]

Running battery tests with an external monitor is not a good way to test because efficiency will depend on the internal circuitry of the connections that drive the external monitor (the port type) and the nature of the display engine of the SoC that drives the monitor.

For example, it will depend on refresh-rate how high/low the memory and other parts will clock when you connect the laptop to an external monitor.

Also, somebody like Apple, which makes products with use cases in mind that involve driving multiple displays, often very high resolution, over DP-on-USB-C/Thunderbolt - they leverage massive display engines that are "slow and wide", which nobody else in the industry does. This allows them to save more power versus the competition when set up identically with multiple displays.

The best way to measure power isolating the internal display is to set its brightness to zero.

All this is from a conversation I had with AT's erstwhile SSD reviewer Billy Tallis on reddit.

Yeah I was thinking about this too recently. Like the display engine is involved here and that’s going to differ depending on the architecture, along with the resolution of the display or even standard involved in compressing the data (and likely an interactive effect there with the display engines).

Didn’t know that for certain about Apple but based on the size of them I figured it’s possible something like that might be going on. I think it’s kind of stupid though — other than for the internal display — to spend too many transistors on marginal area to reduce power here because 99% of the time people will be plugged into a power source when using a monitor or two. Unless power consumption is really just over the top such that the heat laterally dissipated causes meaningful cooling issues or something, it’s ok.

From PHX2 die shots their engines are like 1/2 to 1/3 the size of Zen 4c with SRAM, whereas Apple’s are the size of a P core with a bit of L2.

 

[SiliconFly]

Indeed, and honestly I wonder how many more pages it will take for people to realize and agree that this CPU is simply pathetic with its "centrino moment"...
Idk, lol

Too soon to say. Proper power usage tests are still not available. And the existing tests are all over the place showing massive improvement in power efficiency to no improvement at all. We need updated pcode and better test results as MTL is a new architecture.

 

[SiliconFly]

...their fabric for their main cores isn’t great for idle or low power...

Actually, MTL's NOC fabric is extremely power efficient and far more power efficient than AMD's Infinity fabric. The single culprit for higher power draw in MTL is the CPU tile (RWC) itself. Maybe requires a pcode update or is just plain bad considering it is an ageing design. Either way, it's time for RWC to go away.

 

[TESKATLIPOKA]

Actually, MTL's NOC fabric is extremely power efficient and far more power efficient than AMD's Infinity fabric.

This is based on what?