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.)

 

[TwistedAndy]

It might not outright consume more but it will have comparable peak power draw (all LNL SKUs have a 30W PL2) and significantly lower performance across power profiles. The 17W PL1 benchmarks should be a decent approximation of the M3 in terms of power profile (short burst at 30W for both, dropping to 17/20W sustained) but across the board LNL delivers 20/30% worse performance. At 30W sustained, LNL barely matches M3 in Cinebench.

Where do those 20-30% come from?

If we compare the available Geekbench 6 ST scores, the Lunar Lake ES shows a similar performance to Apple M3 (3000 vs. 2700 + 10% Geekbench OS tax).

 

[SiliconFly]

Where do those 20-30% come from?

If we compare the available Geekbench 6 ST scores, the Lunar Lake ES shows a similar performance to Apple M3 (3000 vs. 2700 + 10% Geekbench OS tax).

Are you sure LNL's GB6 ST score is close to M3???

 

[TwistedAndy]

Are you sure LNL's GB6 ST score is close to M3???

We can simply compare those results:



If we use Linux instead of Windows we will get nearly the same results. Geekbench 6 running in Linux or Unix (macOS) shows 10% better results on the same hardware and frequency. The previous version (Geekbench 5) was more consistent in this regard. The difference was 5-6%.

 

[poke01]

We can simply compare those results:

View attachment 102032

If we use Linux instead of Windows we will get nearly the same results. Geekbench 6 running in Linux or Unix (macOS) shows 10% better results on the same hardware and frequency. The previous version (Geekbench 5) was more consistent in this regard. The difference was 5-6%.

I'm not arguing that. What I'm saying is that Lunar lake needs to clock much higher to even match the M3. I don't know why you don't mention that when you talk efficiency.

Its needs to clock 20% higher than the M3 to achieve the same score so can how lunar be more efficient?

 

[Magio]

Where do those 20-30% come from?

If we compare the available Geekbench 6 ST scores, the Lunar Lake ES shows a similar performance to Apple M3 (3000 vs. 2700 + 10% Geekbench OS tax).

I'm talking about multi-threaded performance, as I said ST is a better showing but with the caveat that platform TDP doesn't tell us much about power consumption in single threaded performance.

https://twitter.com/x/status/1805718733207298541

In both GB 5.5 and CB23 MT, the 17W LNL is 20/30% behind M3.

 

[SiliconFly]

We can simply compare those results:

View attachment 102032

If we use Linux instead of Windows we will get nearly the same results. Geekbench 6 running in Linux or Unix (macOS) shows 10% better results on the same hardware and frequency. The previous version (Geekbench 5) was more consistent in this regard. The difference was 5-6%.

Thats impressive. I'm not sure how significant this is, but LNL GB6 ST score is 10% slower than M3. So, we can say it's very near, but far from equal. Still an impressive feat though!

 

[TwistedAndy]

I'm not arguing that. What I'm saying is that Lunar lake needs to clock much higher to even match the M3. I don't know why you don't mention that when you talk efficiency.

Its needs to clock 20% higher than the M3 to achieve the same score so can how lunar be more efficient?

The clocks alone do not matter that much. The thing that really matters is the actual power consumption.

The P-cores in Apple Silicon were initially designed to be wide. M1 is the 8-wide architecture, M3 is 9-wide, and M4 is 10-wide. This approach requires much more complex structures and much more transistors. This approach offers better performance at lower clock speeds because the frequency/power curve is usually steeper at the beginning.

However, this approach has an obvious limitation. When you want to increase frequency, the power consumption will grow much faster compared to more narrow architectures.

As a result, the power consumption of the P-core in M4 more than doubled compared to M1 (8.95W vs 3.92W in SPEC 2017 FP) while being on the newer node. The frequency has increased from 3.2 to 4.5 GHz. I think the P-core in Apple M5 will reach the 5 GHz barrier and will consume 15W.

For comparison, Intel Lion Cove, Skymont, and Zen 5 are 8-wide architectures, which were initially optimized to run on higher clocks while consuming the reasonable amount of power.

I'm talking about multi-threaded performance, as I said ST is a better showing but with the caveat that platform TDP doesn't tell us much about power consumption in single threaded performance.

You missed the phrase "subject of changes".

In both GB 5.5 and CB23 MT, the 17W LNL is 20/30% behind M3.

In the Geekbench 5 ST test Apple M3 scores between 2200-2300, MT is nearly 10000:

Spoiler: Geekbench 5 Scores for Apple M3

It's a 10-15% difference (2250 - 2000 / 2250 = 11.1%, 10000 - 8800 / 10000 = 12%). It is not 20-30%.

What is more interesting is that those results were obtained for early Pre-QS samples running on 17W PL1 and 17W PL2. Apple M3 is using ~30W PL2 and throttles down to 20-22W under MT tests.

 

[poke01]

For comparison, Intel Lion Cove, Skymont, and Zen 5 are 8-wide architectures, which were initially optimized to run on higher clocks while consuming the reasonable amount of power.

How'd you even know without testing? imo, your bias against Apple is clearly showing. Doesn't Alder Lake, Meteor Lake and Zen4 consume more power than M4 P core in 1t? What does "reasonable amount of power" of power here mean?

As a result, the power consumption of the P-core in M4 more than doubled compared to M1 (8.95W vs 3.92W in SPEC 2017 FP) while being on the newer node. The frequency has increased from 3.2 to 4.5 GHz.

That's what happens when you increase clocks, pwr increases. But at the same clocks( at 4Ghz) M4 power barely consumption increased while performance increased. If power consumption increased at the same clocks then its a problem, but that's not the case with M4.


At 4.5GHz, yes power consumption increased but you also get more performance.

However, this approach has an obvious limitation. When you want to increase frequency, the power consumption will grow much faster compared to more narrow architectures.

Not true. This also happens with Intel and AMD.

 

[SiliconFly]

The clocks alone do not matter that much. The thing that really matters is the actual power consumption.

The P-cores in Apple Silicon were initially designed to be wide. M1 is the 8-wide architecture, M3 is 9-wide, and M4 is 10-wide. This approach requires much more complex structures and much more transistors. This approach offers better performance at lower clock speeds because the frequency/power curve is usually steeper at the beginning.

However, this approach has an obvious limitation. When you want to increase frequency, the power consumption will grow much faster compared to more narrow architectures.

As a result, the power consumption of the P-core in M4 more than doubled compared to M1 (8.95W vs 3.92W in SPEC 2017 FP) while being on the newer node. The frequency has increased from 3.2 to 4.5 GHz. I think the P-core in Apple M5 will reach the 5 GHz barrier and will consume 15W.

For comparison, Intel Lion Cove, Skymont, and Zen 5 are 8-wide architectures, which were initially optimized to run on higher clocks while consuming the reasonable amount of power.



You missed the phrase "subject of changes".



In the Geekbench 5 ST test Apple M3 scores between 2200-2300, MT is nearly 10000:

Spoiler: Geekbench 5 Scores for Apple M3

View attachment 102033

It's a 10-15% difference (2250 - 2000 / 2250 = 11.1%, 10000 - 8800 / 10000 = 12%). It is not 20-30%.

What is more interesting is that those results were obtained for early Pre-QS samples running on 17W PL1 and 17W PL2. Apple M3 is using ~30W PL2 and throttles down to 20-22W under MT tests.

Excellent write-up!

(1) if M5's single P core touched 15W, Apple consumers will go berserk!

(2) I'm still coming to terms with LNL being close to M3 (with additional headroom due to PL2).

 

[poke01]

I'm still coming to terms with LNL being close to M3 (with additional headroom due to PL2).

since this is a ES, retail Lunar will be even better. Lunar already has a better software support than Apple CPUs. All my games work with Lunar. What I don't like is setting false expectations regarding battery life vs M3, lets wait for the tests but its looking great so far.

 

[xiewe3wq]

(1) if M5's single P core touched 15W, Apple consumers will go berserk!

Why would Apple customers care about power measurements? They care about real life battery life and performance, both which are excellent with M4.

 

[Hitman928]

And how many dirt cheap processors can actually do a respectable job of replacing human labor for advise at desk? A kiosk with AI tuned for the specific application is nothing cost wise in comparison.

I would expect they would probably just connect to a server where the processing happens if they need to run some LLM.

 

[TwistedAndy]

How'd you even know without testing? imo, your bias against Apple is clearly showing. Doesn't Alder Lake, Meteor Lake and Zen4 consume more power than M4 P core in 1t? What does "reasonable amount of power" of power here mean?

Let's do some napkin math.

Apple M3 runs that test with nearly a 28-30W power limit (first run, ~10000-10500 points in GB 5 MT) and then throttles to 20W (~8000) or even lower - to 12W (~6000).

That's why we have so different results for MacBook Air with M3 in Geekbanch 5 database. They vary from 6000 to 10500 points. You can find them here.

According to Geekerwan's charts, M3 P-core at 3.8 GHz consumes 5.3W. If we assume that all the cores are working on that frequency, the whole SoC will consume 5.3*4 + 0.75*4 = 24W + memory and other stuff = 28W. That's what we have during the first run. (Geekerwan has more precise measurements for E-cores, but I'm too lazy to find that).

In the case of Lunar Lake, P-cores probably consume nearly 3W in MT load, and the E-cores - nearly 0.75W. 3 * 4 + 4 * 0.75 = 15W + memory and other stuff = 17W.

If we compare Apple M3 running on 17W PL1 with the pre-QS sample of Lunar Lake running on the same 17W PL1, the Lunar Lake will be 10% faster than M3 (~8000 vs ~8800).

That's what happens when you increase clocks, pwr increases. But at the same clocks( at 4Ghz) M4 power barely consumption increased while performance increased. If power consumption increased at the same clocks then its a problem, but that's not the case with M4.

We got 7% more performance for 4% more power on the newer node. That's fine.

 

[dullard]

I would expect they would probably just connect to a server where the processing happens if they need to run some LLM.

Why on earth would a kiosk run an LLM? I cannot see a single reason that an info desk at an airport needs to be trained on how to plumb your kitchen sink. And vise versa, why would a kiosk at your local hardware store need to be trained on airline routes to Kazakhstan? Large language models have no purpose there. Think small language models specifically tailored to the purpose.

 

[poke01]

Why would airports kiosks even need lunar lake????

Lunar is made for client not commercial. It’s not economical. TSMC N3B isn’t cheap

 

[poke01]

In the case of Lunar Lake, P-cores probably consume nearly 3W in MT load, and the E-cores - nearly 0.75W. 3 * 4 + 4 * 0.75 = 15W + memory and other stuff = 17W.

If that is the case then Intel won 2024

 

[dullard]

what software do you use?

I'm in biotech. I work with data from dozens of instruments: none of which have Apple software. I have 25 years of the software that I programmed to analyze the data--I'm certainly not taking the time to convert those over. Then more mundane things like Solidworks crash and/or have really poor emulation.

 

[dullard]

Why would airports kiosks even need lunar lake????

Lunar is made for client not commercial. It’s not economical. TSMC N3B isn’t cheap

To answer verbal questions relevant to your trip.

Oh and tell that to the kiosk manufacturers using Intel chips.
https://www.digitalsignagetoday.com/press-releases/embrace-the-era-of-ai-pc-together-with-giada/

In 2024, Giada will release its new AI PC adopting Intel Core Ultra processors

https://sodaclick.com/pages/hands-free-solutions/
etc.

I expect we'll here more, including Lunar Lake specific products, at the next AI conference.

 

[poke01]

I'm in biotech. I work with data from dozens of instruments: none of which have Apple software. I have 25 years of the software that I programmed to analyze the data--I'm certainly not taking the time to convert those over. Then more mundane things like Solidworks crash and/or have really poor emulation.

That sounds interesting. Thanks for sharing.

 

[Magio]

https://twitter.com/x/status/1806396250809335831

Something I'd missed which does change some things, the scores published by this account as being 17W are for a chip with PL1 and PL2 both being 17W.

The 30W numbers are a bit weird right now so it's tough to predict how a 17W PL1/30W PL2 chip (which is all SKUs but one) will behave, but it's a positive for the scores that were a bit underwhelming and super encouraging for those that were already good (for example TimeSpy GPU). Putting the scores in spoiler for context for those who missed the tweet.

Spoiler

https://twitter.com/x/status/1805718733207298541

 

[dttprofessor]

Let's do some napkin math.

Apple M3 runs that test with nearly a 28-30W power limit (first run, ~10000-10500 points in GB 5 MT) and then throttles to 20W (~8000) or even lower - to 12W (~6000).

That's why we have so different results for MacBook Air with M3 in Geekbanch 5 database. They vary from 6000 to 10500 points. You can find them here.

According to Geekerwan's charts, M3 P-core at 3.8 GHz consumes 5.3W. If we assume that all the cores are working on that frequency, the whole SoC will consume 5.3*4 + 0.75*4 = 24W + memory and other stuff = 28W. That's what we have during the first run. (Geekerwan has more precise measurements for E-cores, but I'm too lazy to find that).

In the case of Lunar Lake, P-cores probably consume nearly 3W in MT load, and the E-cores - nearly 0.75W. 3 * 4 + 4 * 0.75 = 15W + memory and other stuff = 17W.

If we compare Apple M3 running on 17W PL1 with the pre-QS sample of Lunar Lake running on the same 17W PL1, the Lunar Lake will be 10% faster than M3 (~8000 vs ~8800).



We got 7% more performance for 4% more power on the newer node. That's fine.

MTL 2 LPE crestmont 2.2GH
LNL 4 LPE skymont 3.7GH
Intel should have different strategy，prefer to 4 skymont cores (maybe 3.0GH max all cores) and shut down 4 arrow lake cores on the ring for battery life.

 

[TwistedAndy]

Intel should have different strategy，prefer to 4 skymont cores (maybe 3.0GH max all cores) and shut down 4 arrow lake cores on the ring.

That's the main reason why Intel has moved E-cores to a separate island, added SLC (side cache), and increased the performance of those E-cores. Intel wants to keep the P-cores and related stuff turned off for light loads.

 

[Hitman928]

To answer verbal questions relevant to your trip.

Oh and tell that to the kiosk manufacturers using Intel chips.
https://www.digitalsignagetoday.com/press-releases/embrace-the-era-of-ai-pc-together-with-giada/

https://sodaclick.com/pages/hands-free-solutions/
etc.

I expect we'll here more, including Lunar Lake specific products, at the next AI conference.

There might be some niche ones that do more, but most everything you see in the general public space is very cheap hardware for customer interfacing. They want things that are low cost, reliable, and very easy to fix if necessary. I'm sure there are companies trying to make things happen, but that doesn't mean it will actually achieve market adoption any time soon. Based on history, when it does happen, it will be because there are cheap, reliable solutions that target the use case.

Maybe you're right, we'll see, but I would be shocked if there is anything outside of maybe of foot note of results from companies buying LNL to drive AI kiosks.

 

[SiliconFly]

Brett @ UFD Tech says 30W Lunar Lake beats the latest 54W Ryzen AI 9 HX in graphics!

Doesn't add up! Or is Battlemage that good?

 

[Saylick]

Brett @ UFD Tech says 30W Lunar Lake beats the latest 54W Ryzen AI 9 HX in graphics!

Doesn't add up! Or is Battlemage that good?

Allegedly, the 890M is bandwidth starved, so it doesn’t scale much higher than it’s performance at 30W, even if you give it more juice.