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.)

 

[mikk]

SPECint ST efficiency comparison

https://twitter.com/x/status/1741390546562883694

View attachment 91238

At 10W slightly over 20% more efficient than Raptor Lake and at 5W slightly over 30%. This is what we should get from from Meteor Lake in a typical power envelope in multithreaded benchmarks and I think we already see this in some of the newer reviews. Not as good as Phoenix but this gap is not the end of the world.

 

[Hulk]

SPECint ST efficiency comparison

https://twitter.com/x/status/1741390546562883694

View attachment 91238

The fact that the E cores on MTL in this graph are tapping out so early is making me think these are the LPE's and not the E's on the CPU tile. I would have thought (hoped) Crestmont would be more energy efficient than Gracemont but I don't want to make any implications regarding that from this graph with only 4 data points from the mysterious MTL E's.

One thing is certain, the 7840U is a monster when it comes to efficiency. I wonder how much of the 7840U advantage is due to architecture and how much is due to process?

Looks like Intel did get a nice bump in efficiency with Intel 4 vs Intel 7 especially when you consider MTL gives up some efficiency (when the CPU tile is active) to RPL due to RPL's monolithic design. While Intel is still behind in mobile I am ready to say that MTL is a solid "hit" for Intel, they got on base. They've reset their field with MTL, meaning they have a new base line of performance and efficiency after making the difficult move to tiles. For a first go this isn't a home run but they didn't strike out either. I'd say it's a double in baseball terms.

I wish the 7840U was available in more mobile devices. It's a phenom.

 

[Abwx]

At 10W slightly over 20% more efficient than Raptor Lake and at 5W slightly over 30%. This is what we should get from from Meteor Lake in a typical power envelope in multithreaded benchmarks and I think we already see this in some of the newer reviews. Not as good as Phoenix but this gap is not the end of the world.

You should look at the horizontal difference, power used to process a task in the same time, that s efficency comparison, so take MTL at say 10W and look how much power the 7840U use to perform at the same level.

7840U use 7W to perform like the 155H at 10W, that s 43% better efficency at this performance point.

 

[mikk]

You should look at the horizontal difference, power used to process a task in the same time, that s efficency comparison, so take MTL at say 10W and look how much power the 7840U use to perform at the same level.

7840U use 7W to perform like the 155H at 10W, that s 43% better efficency at this performance point.

This is not how it works, you have to test the efficiency or performance at the same power.

 

[maddie]

This is not how it works, you have to test the efficiency or performance at the same power.

Not same workload? That's the standard for other fields.

 

[Hitman928]

Not same workload? That's the standard for other fields.

On the consumer side, efficiency is typically measured as performance at the same power or performance at the default power (whatever the processor uses by default). This is because power cost is typically a lower concern on the user side. Users are typically more concerned with performance and efficiency during workloads without a set amount of work to finish (i.e., gaming, web browsing, etc.). For efficiency, the main concern of most users is the upfront cost/complexity of the cooling system and maybe how much heat will be dumped into their room during use. All of this is captured well with measuring performance at the same power or at the default power. Laptops sometimes are the exception to this because of battery life concerns but battery life is a complex measurement that involves much more than just the CPU and if a user is really doing something compute intensive, will know that running plugged in for anything but short runs will be needed no matter what.

Where power used at the same performance (really an energy used calculation is the appropriate measurement) comes more into play is usually on the server side where energy costs are a large consideration. This can be especially true with cloud setups where companies are paying for server time, which cost per minute is tied to some degree to power consumed. In such cases, measuring power for a given performance level makes a lot of sense.

 

[Abwx]

Not same workload? That's the standard for other fields.

AFAIK two cars are compared for a same distance at a same speed, if one car has to do the same distance within 0.8x the time it will require 25% higher speed, and since the required energy increase as the square of speed it will consume 56% more.

Likewise with CPUs power also increase as a square of speed, so if a CPU is required to do more work within a same time it will be inherently gimped perf/watt wise the same as in the exemple above...

 

[TESKATLIPOKA]

This is not how it works, you have to test the efficiency or performance at the same power.

Who said you have to test It like you said?
You either test It at ISO performance, which @Abwx did, or ISO power like you did.
You want to use ISO power, because the difference is smaller.
@Abwx wants to use ISO performance because the difference is bigger.
You want to make MTL look better, he wants It to look worse, yet both of you are correct.

 

[Abwx]

Who said you have to test It like you said?
You either test It at ISO performance, which @Abwx did, or ISO power like you did.
You want to use ISO power, because the difference is smaller.
@Abwx wants to use ISO performance because the difference is bigger.
You want to make MTL look better, he wants It to look worse, yet both of you are correct.

You are wrong in this one, i have no other intention than explaning watt is a perf/watt comparison, basically that s comparing the energy taken to do a given amount of work within a same time.

Anything other than this definition is not perf/watt but another metric that is confused with perf/watt.

FI a new process is said to provide the same perf at 0.5x the power, so that s 2x the perf/watt, that is, 2x less power is required to do the same amount of work within a same time.

Now the same process speed improvement is about 1.4x, so perf of the new process is 1.4x higher, that is , it s another metric, it s speed, aka frequency, improvement at a same power, or said another way you can do 1.4x the work at same power and same time.

 

[TESKATLIPOKA]

You are wrong in this one, i have no other intention than explaning watt is a perf/watt comparison, basically that s comparing the energy taken to do a given amount of work within a same time.

Anything other than this definition is not perf/watt but another metric that is confused with perf/watt.

Are you sure?
You wrote this:

7840U use 7W to perform like the 155H at 10W, that s 43% better efficency at this performance point.

Does this look like perf/W to you? To me It looks more like W/perf.

Performance is the time needed to finish a given work. Let's say It takes 1hour.
So 7840U would consume 7W to finish It, but 155H will consume 10W during this hour.
7840U has 30% better power efficiency or 155H has 43% worse power efficiency.

 

[Abwx]

Performance is the time needed to finish a given work. Let's say It takes 1hour.
So 7840U would consume 7W to finish It, but 155H will consume 10W during this hour.
7840U has 30% better power efficiency or 155H has 43% worse power efficiency.

It s the other way around, 7840U has 43% better perf/watt and 155H has 30% worse power efficency...

Or said otherwise the 7840U does 43% more work for a same amount of energy and the 155H does 30% less work for said energy amount.

 

[cebri1]

What's the perf/watt advantage at 15W?

 

[TESKATLIPOKA]

It s the other way around, 7840U has 43% better perf/watt and 155H has 30% worse power efficency...
Or said otherwise the 7840U does 43% more work for a same amount of energy and the 155H does 30% less work for said energy amount.

This is simply not correct.

If you check that graph, what do you see?
At 7W I see 7840U managing ~6.9 while 155H managed only ~5.9, that's 17% better perf/W.
At 10W I see 7840U managing ~8.1 while 155H managed only ~7.4, that's 9.5% better perf/W.
There is clearly no 43% more work done or 43% better perf/W how you calculated It.
As I previously said, you didn't calculate perf/W but W/perf instead.

P.S. Also what I wrote originally that 7840U has 30% better power efficiency(W/perf) or 155H has 43% worse power efficiency(W/perf) is correct.
Because 7W vs 10W, or 30% lower in this case smaller number is better.

 

[Saylick]

The fact that the E cores on MTL in this graph are tapping out so early is making me think these are the LPE's and not the E's on the CPU tile. I would have thought (hoped) Crestmont would be more energy efficient than Gracemont but I don't want to make any implications regarding that from this graph with only 4 data points from the mysterious MTL E's.

FWIW, while the E core's perf/W curve is rather lack luster, if you normalized by the area of the core (mm2), that E core curve is rather competitive with Zen 4 mobile if one assumes that Z4 is twice as large as Crestmont. Of course, it will be competitive for MT applications only.

 

[Abwx]

This is simply not correct.

If you check that graph, what do you see?
At 7W I see 7840U managing ~6.9 while 155H managed only ~5.9, that's 17% better perf/W.

How much would the 155H require to do 6.9, that is, to complete the task in the same time as the 7840U..?.

Or eventually if you downclock the 7840U such as it score 5.9 and hence set such that it use the same time to perform the Spec-int task that is used for this bench.

Because here the 7840U is required to complete the task 17% faster, at this rate if a 6 + 8 RPL is set such that it use 3x the time to execute the task it will be way more efficient than any of those two CPUs.

 

[TESKATLIPOKA]

How much would the 155H require to do 6.9, that is, to complete the task in the same time as the 7840U..?.

Or eventually if you downclock the 7840U such as it score 5.9 and hence set such that it use the same time to perform the Spec-int task that is used for this bench.

Because here the 7840U is required to complete the task 17% faster, at this rate if a 6 + 8 RPL is set such that it use 3x the time to execute the task it will be way more efficient than any of those two CPUs.

As I said, you are talking about W/perf(ISO Performance) yet you are calling It as perf/W(ISO power).

 

[H433x0n]

Looks like Intel’s graphic of Intel 4 vs Intel 7 was actually accurate.

 

[Abwx]

As I said, you are talking about W/perf(ISO Performance) yet you are calling It as perf/W(ISO power).

Yes, because if you dont take account of the time those numbers do not make sense, as said RPL clocked low enough would have better perf/watt than both MTL and 7840U, but it would use much more time to complete a same task.

 

[H433x0n]

The fact that the E cores on MTL in this graph are tapping out so early is making me think these are the LPE's and not the E's on the CPU tile. I would have thought (hoped) Crestmont would be more energy efficient than Gracemont but I don't want to make any implications regarding that from this graph with only 4 data points from the mysterious MTL E's.

One thing is certain, the 7840U is a monster when it comes to efficiency. I wonder how much of the 7840U advantage is due to architecture and how much is due to process?

Looks like Intel did get a nice bump in efficiency with Intel 4 vs Intel 7 especially when you consider MTL gives up some efficiency (when the CPU tile is active) to RPL due to RPL's monolithic design. While Intel is still behind in mobile I am ready to say that MTL is a solid "hit" for Intel, they got on base. They've reset their field with MTL, meaning they have a new base line of performance and efficiency after making the difficult move to tiles. For a first go this isn't a home run but they didn't strike out either. I'd say it's a double in baseball terms.

I wish the 7840U was available in more mobile devices. It's a phenom.

I don’t think we’re getting an accurate picture with the e-cores. All that’s listed here is the lp ecores that are just more dense versions of Gracemont. I also hope that the P cores were running totally independently of the lp ecores otherwise this data is erroneous since it’d have a ton of uncore sapping the power.

The Crestmont cores should have the same if not better perf/watt uplift over Redwood Cove. It’s got the same node uplift and the advantage of +5% IPC.

 

[Khato]

View attachment 91238

Nice find... Though I'm taking the absolute value of the 'package power' numbers with a grain of salt. What all is included in that figure in order to make it start in the 4-5W range for each of the curves? If it is just CPU package power, then there's no way that package power would start at 5W when just the 2 LP E-cores are active. I'd guess that the actual core power for the first point on each of the curves is probably around 0.5W?

 

[Hulk]

I don’t think we’re getting an accurate picture with the e-cores. All that’s listed here is the lp ecores that are just more dense versions of Gracemont. I also hope that the P cores were running totally independently of the lp ecores otherwise this data is erroneous since it’d have a ton of uncore sapping the power.

The Crestmont cores should have the same if not better perf/watt uplift over Redwood Cove. It’s got the same node uplift and the advantage of +5% IPC.

I agree. The E cores are notoriously difficult to isolate for testing. Hard data on MTL is still MIA.

 

[Hitman928]

You should look at the horizontal difference, power used to process a task in the same time, that s efficency comparison, so take MTL at say 10W and look how much power the 7840U use to perform at the same level.

7840U use 7W to perform like the 155H at 10W, that s 43% better efficency at this performance point.

So using your methodology and the data given, how much more efficient is the M1 compared to the 7840u?

 

[H433x0n]

So using your methodology and the data given, how much more efficient is the M1 compared to the 7840u?

This should be good .. got to get my🍿ready.

 

[Abwx]

So using your methodology and the data given, how much more efficient is the M1 compared to the 7840u?

Apply the same methodology, to evaluate what are the respective joules when performing a same work in a same time, dunno the exact number but according to NBC the M1 is ahead but the 7840U is not that far behind.

Edit : According to NBC the M1 score the same in CB R23 as a 7840U that is at about 30W, i dont think that the M1 is that much below 30W.

Processor Comparison - Head 2 Head

AMD Ryzen 7 PRO 7840U vs Apple M1 vs Apple M1 Pro - Benchmarks, Tests and Comparisons

www.notebookcheck.net

 

[Hitman928]

Apply the same methodology, to evaluate what are the respective joules when performing a same work in a same time, dunno the exact number but according to NBC the M1 is ahead but the 7840U is not that far behind.

Why do you say you don’t know the number? I’m asking you to use your method to calculate the efficiency difference.

Edit: I'm not talking about CB R23, I'm talking about for Spec_int.