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	12 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)	Lunar Lake	Panther Lake
Platform	Mobile H/U Only	Desktop Only	Desktop & Mobile H&HX	Mobile U Only	Mobile H
Process Node	Intel 4	Intel 20A	TSMC N3B	TSMC N3B	Intel 18A
Date	Q4 2023	Q1 2025 ?	Desktop-Q4-2024 H&HX-Q1-2025	Q4 2024	Q1 2026 ?
Full Die	6P + 8P	6P + 8E ?	8P + 16E	4P + 4E	4P + 8E
LLC	24 MB	24 MB ?	36 MB ?	12 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.)

 

[Hulk]

The more I think about that LL 10k CB R23 MT score the more confused I am.

Assuming 4.5/3.5 clocks and Lion Cove performs 5% better than Raptor Cove in this bench, and Skymont perform performs 5% worse than Raptor Cove then the resulting score should be nearly 15k.

Either Intel's entire LL presentation, all of the engineer talks, the slides, all of it, was gaslighting, or that leak is faulty.

We shall know soon enough!

 

[DavidC1]

The more I think about that LL 10k CB R23 MT score the more confused I am.

Assuming 4.5/3.5 clocks and Lion Cove performs 5% better than Raptor Cove in this bench, and Skymont perform performs 5% worse than Raptor Cove then the resulting score should be nearly 15k.

They are early tests, user submitted tests, plus very power limited, which is what MT is.

You'll get decent ST performance approximations but not for MT. There's no way 4 Lion Cove cores alone will clock at 4.5GHz and fit at 17W, nevermind 4 Skymont on top of it. Even at 30W it'll be a challenge for 4P @ 4.5GHz.

30W package power means 25W for core which means each core has only 6.25W available to it. What can Lion Cove do at 6.25W? What can Golden Cove do at 6.25W?

Here's for comparison Meteorlake:

30W: 10,000
17W: ~6,500

The 125H requires ~70W to reach not even 15K. Even the 7840HS requires 50W.

If the Xitter leak is true where performance increases at 30W is not much then Lunarlake has a steeper V/F curve where it is even more optimized for lower frequency efficiency.

@majord Ok. So AMD must have followed Intel and did opposite on the later chips to favor high frequency.

Lunarlake is still about leadership battery life. It has everything a regular user requires. High ST performance for responsiveness, great graphics, good battery life, and potentially small form factor.

 

[DavidC1]

Something little different.

@Abwx commonly makes the argument that since frequency and power relationship is cube root squared, that's how the power and frequency graphs should always follow.

This is NOT the case. Every design has an optimal operating frequency. You'll know if you played extensively with V/F graphs, usually easiest when done with GPUs.

I have. The Polaris clocks are nearly at the floor at 0.6V. Sure, 0.6V would require less power than 0.7V, but that 100mV increase means suddenly your max clocks can go from 300MHz to 600MHz.

And on the peak, you can undervolt it to 950mV and have the GPU clock at more than 2GHz. So there's a linear portion of the graph and exponential portion of the graph in regards to frequency. On the exponential part of the graph you can give just 5% extra voltage and have it clock 10, 20, or even 30% higher in clocks!

So rather than 1.3x clocks being 1.3x1.3x1.3x or 2.2x in power, it's 1.3x1.05x1.05 or just a 43% increase in power, likely less than that because there's static power contribution and voltages are just about active power. Static power contribution is not negligible. It is in the order of 10-20%.

 

[Abwx]

Something little different.

@Abwx commonly makes the argument that since frequency and power relationship is cube root squared

I never said cube root squared but around 2.36 - 2.56 root squared for usual TSMC s processes, for LNL it is :

CB score ratio = 10200/8200 = 1.244

Power ratio is 28/15 = 1.87

P/f evolution exponent is hence ln(1.87)/ln(1.244) = 2.867

So that yield P(f) = f^2.867 and conversly F(p) = P^(1/2.867)

So more precisely 2.867 root squared in this TDP range, wich is signifcantly more than previous TSMC s processes, and indeed close to a cubic root, but that s an exception so far, and so much that it s unlikely that it s due to the process but rather to the hybrid uarch that is not optimal if the e and P cores V/f curves are significantly different, or eventually that at 17W there s a turbo at work wich is not present at 30W.

 

[DavidC1]

I never said cube root squared but around 2.36 - 2.56 root squared for usual TSMC s processes, for LNL it is :

The problem with your analysis is the same AMD/Intel makes in their presentations, which is being oddly accurate like claiming "17%" gains. In the sane days we used to call it 15-20%.

You are assuming the early platform is completely representative of the final product, and the testing is done in a perfect environment. That few % will skew the numbers to say one is winning over the other since we're arguing about mere single digit percentages.

The natural variation present in user-submitted tests, in a prototype platform, from one tester, likely time constrained at that too, and with limited information presented to us, and on a laptop no less will render it no more than a case of "oh that's interesting" and a reference point than "This is how the product performs".

 

[Abwx]

The problem with your analysis is the same AMD/Intel makes in their presentations, which is being oddly accurate like claiming "17%" gains. In the sane days we used to call it 15-20%.

You are assuming the early platform is completely representative of the final product, and the testing is done in a perfect environment. That few % will skew the numbers to say one is winning over the other since we're arguing about mere single digit percentages.

There are another leaked number to substantiate LNL perfs at 30W.

ARL s 8 + 16 score 43118 pts at 250W and 5.4 + 4.6 frequencies.

A 4 + 8 would score around 21600 at 125W and same frequencies.

A 2 + 4 would score 10800 at 62.5W., also at 5.4 + 4.6.

With 33 % more cores, and hence 1.33x lower frequency, you can get those 10800 pts at 1.33^2.8 = 2.22x lower power or only 2x if the exponent is less than 2.8.

So that makes 62.5/2.22 = 28W for 10800 pts or eventually 62.5/2 = 31W, the little 600 pts difference is due to the uncore that wont scale exactly as much as the core powers or to anything else, but you ll admit than using two different designs we end standing on the same feet.

 

[majord]

They are early tests, user submitted tests, plus very power limited, which is what MT is.

You'll get decent ST performance approximations but not for MT. There's no way 4 Lion Cove cores alone will clock at 4.5GHz and fit at 17W, nevermind 4 Skymont on top of it. Even at 30W it'll be a challenge for 4P @ 4.5GHz.

30W package power means 25W for core which means each core has only 6.25W available to it. What can Lion Cove do at 6.25W? What can Golden Cove do at 6.25W?

Here's for comparison Meteorlake:

30W: 10,000
17W: ~6,500

The 125H requires ~70W to reach not even 15K. Even the 7840HS requires 50W.

If the Xitter leak is true where performance increases at 30W is not much then Lunarlake has a steeper V/F curve where it is even more optimized for lower frequency efficiency.

@majord Ok. So AMD must have followed Intel and did opposite on the later chips to favor high frequency.

Lunarlake is still about leadership battery life. It has everything a regular user requires. High ST performance for responsiveness, great graphics, good battery life, and potentially small form factor.

Yeah I'm not sure what is going on with that 7840HS dropping off, but as alluded to, perhaps the VF curve is whack at lower TDPs. Check U series performs far better at that 15-17w .

Re Leadership in battery life, well that is always welcome, but I'm not sure how relevant it will be in actual devices outside perhaps pure idle time. Assuming it can dip down to Snapdragon levels of idle power.

At the end of the day the device has a far bigger bearing on actual life when talking about laptops with any of the modern SoCs. Any advantage in idle and light use consumption can get completely smeared away by power hungry OLEDs , and poor consumption of other peripherals. Therefore It's hard to sell an SoC on that alone when it's assessed as part of a device

 

[511]

At the end of the day the device has a far bigger bearing on actual life when talking about laptops with any of the modern SoCs. Any advantage in idle and light use consumption can get completely smeared away by power hungry OLEDs , and poor consumption of other peripherals. Therefore It's hard to sell an SoC on that alone when it's assessed as part of a device

I hope they rework Intel EVO spec for LNL EVO was a nice idea but not well executed

 

[DavidC1]

It is inaccurate to say Lunarlake is a 15W chip + 2W DRAM. DRAM peak is harder to utilize in real world applications than the chip, thus it's basically "cheating" to eek out more performance by using slight bit of extra headroom.

My Kabylake Yoga uses 0.7W in average for the DRAM. It goes as low as 0.5W. So if 17W LNL uses 0.5W in average for the DRAM, the CPU is using 16.5W.

Any advantage in idle and light use consumption can get completely smeared away by power hungry OLEDs , and poor consumption of other peripherals. Therefore It's hard to sell an SoC on that alone when it's assessed as part of a device.

Icelake has great idle power use. Very low. The problem is it sucks on web browsing and video playback, so any load makes the low idle useless. The x86 platforms have a problem in both low idle and/or keeping it low.

Luckily the demos they have been doing is in load scenarios as well. 40% reduction in teams, 40% reduction in video playback and frame-rate capped gaming. Interestingly, they quote that 40% number for memory PHY power reduction.

Lunarlake using similar power doing local video playback as Meteorlake on idle.

 

[511]

It is inaccurate to say Lunarlake is a 15W chip + 2W DRAM. DRAM peak is harder to utilize in real world applications than the chip, thus it's basically "cheating" to eek out more performance by using slight bit of extra headroom.

This 17W is total distributed power budget among CPU+GPU+NPU+RAM+Uncore

 

[DavidC1]

This 17W is total distributed power budget among CPU+GPU+NPU+RAM+Uncore

It don't matter. Uncore and RAM can't go to zero, but the rest can be power gated and essentially be zero, thus CPU/GPU can use all of it. 2W for Lunarlake is misleading. It'll need something like a memory power virus.

Anyways, just had to restart the WMI service on my Kabylake Yoga to get package power down by 0.5W. 10-15% battery life impact thanks to one service.

Thanks Microsoft!!

 

[511]

It don't matter. Uncore and RAM can't go to zero, but the rest can be power gated and essentially be zero, thus CPU/GPU can use all of it. 2W for Lunarlake is misleading. It'll need something like a memory power virus.

Yeah 2W is misleading unless you run prime95

Anyways, just had to restart the WMI service on my Kabylake Yoga to get package power down by 0.5W. 10-15% battery life impact thanks to one service.

Thanks Microsoft!!

I am with you on this one i had i5-8250U for like 3.5 years was very nice but Microsoft services are amazing at screwing modern hardware looks at modern standby

 

[Magio]

Is this actually in mW? 1.4W package power for 4k30 AV1 sounds quite impressively low to me, but then again 14W would be wayyy too high.

There's this video where they demo LNL in this exact scenario and total system power (with screen off) is 5/6W, that seems higher than what I'd expect if the package power on its own was 1.4W.

 

[511]

Is this actually in mW? 1.4W package power for 4k30 AV1 sounds quite impressively low to me, but then again 14W would be wayyy too high.

There's this video where they demo LNL in this exact scenario and total system power (with screen off) is 5/6W, that seems higher than what I'd expect if the package power on its own was 1.4W.

Found it here they could have made improvements considering 3-4 months from launch

 

[Magio]

Found it here they could have made improvements considering 3-4 months from launch

Ah, hadn't seen that video ! 2.5/3W package power on VVC does make AV1 at 1.4W seem plausible !

Do hope we'll see laptops that do a better job keeping system power in check than in that AV1 demo.

 

[511]

The more I think about that LL 10k CB R23 MT score the more confused I am.

Assuming 4.5/3.5 clocks and Lion Cove performs 5% better than Raptor Cove in this bench, and Skymont perform performs 5% worse than Raptor Cove then the resulting score should be nearly 15k.

Either Intel's entire LL presentation, all of the engineer talks, the slides, all of it, was gaslighting, or that leak is faulty.

We shall know soon enough!

Ain't no way it can do 4.5 ghz P and 3.5 E All core

 

[Hulk]

Ain't no way it can do 4.5 ghz P and 3.5 E All core

I agree those frequencies are optimistic for mobile.
Probably more like 4/3, right?

4/3 with Lion Cove doing 5% better than Raptor and Skymont 5% less than Raptor (in CB R23) would show a score of about 13,500.

 

[Bouowmx]

From Hot Chips: Lunar Lake optimizes the glue

 

[igor_kavinski]

Intel Lunar Lake for AI PCs at Hot Chips 2024

In its presentation at Hot Chips 2024, we learned more about the upcoming Intel Lunar Lake SoC for AI PCs and some of the updates to the chip

www.servethehome.com

 

[Hitman928]

View attachment 106248
From Hot Chips: Lunar Lake optimizes the glue

LPE cores on MTL were cache starved and on a different die, so this is not surprising. Comparing to MTL E core would be more interesting, IMO.

 

[dullard]

Intel Lunar Lake for AI PCs at Hot Chips 2024

In its presentation at Hot Chips 2024, we learned more about the upcoming Intel Lunar Lake SoC for AI PCs and some of the updates to the chip

www.servethehome.com

Thanks, I was wondering why none of the sites I visit were giving Hot Chips updates.

 

[DavidC1]

I agree those frequencies are optimistic for mobile.
Probably more like 4/3, right?

4/3 with Lion Cove doing 5% better than Raptor and Skymont 5% less than Raptor (in CB R23) would show a score of about 13,500.

Hulk, Meteorlake gets 6.5K at 17W. You are being way too optimistic here. 15K using 70W. It requires 30W to get 10K. And 125H is a 4+8 part with Hyperthreading meaning it should require less clocks to reach the same performance.

Skymont should do better on lower frequencies but It should be in the low 3GHz frequency for both cores. Maybe even high 2GHz for E.

Is this actually in mW? 1.4W package power for 4k30 AV1 sounds quite impressively low to me, but then again 14W would be wayyy too high.

There's this video where they demo LNL in this exact scenario and total system power (with screen off) is 5/6W, that seems higher than what I'd expect if the package power on its own was 1.4W.

It is. The web browsing power for Meteorlake is not too dissimilar for my Kabylake system.

I would never get an OLED and resolution beyond FHD for my system. Battery life impact is forever, while the Wow! factor from a screen lasts at most a month.

 

[DavidC1]

Intel Lunar Lake for AI PCs at Hot Chips 2024

In its presentation at Hot Chips 2024, we learned more about the upcoming Intel Lunar Lake SoC for AI PCs and some of the updates to the chip

www.servethehome.com

Intel summary:
->20% MT perf with lower core count
->20% ST performance, or same performance as MTL at 1/2 the power
-More than twice the performance per watt in productivity(?)
-LNL reduces SoC power by 40%

Interesting bit:

Intel says removing Hyper-Threading helped them get more efficient at low power.

 

[dullard]

There's this video where they demo LNL in this exact scenario and total system power (with screen off) is 5/6W, that seems higher than what I'd expect if the package power on its own was 1.4W.

Doesn't Wi-Fi 7 take a few watts of power on its own?

 

[DavidC1]

Doesn't Wi-Fi 7 take a few watts of power on its own?

5-6W with screen off isn't particularly efficient. If WiFi 7 takes that much power I don't think it's worth it. Based on those numbers alone it's no advancement at all from my Kabylake on the power front. But Meteorlake system shows nearly double so we'll see.