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.)

 

[cebri1]

Incase we have any photoshop enthusiasts would love to see MTL plotted on this graph. For some reason the OP didn’t include it.

Paint enthusiast here. Line too thick but you get the idea.

 

[Abwx]

The curve of the 7840HS is total non sense in respect of the 7840U, that s just made up numbers.

 

[TESKATLIPOKA]

At 7.5W RPL does 0.65 pts per watt, at 16W the 155H does 0.51 pts per watt, the former is more efficient in this case, that s what i said in another post, if you increase the execution time of a CPU you increase its perf/watt.

In this case it s not a comparison at same throughput, wich is the bias i was talking about, comparing CPUs at different power is not indicative of their respective perf/watt.

You are choosing data points which will show you a result which you want to see.
I can also choose ~2.4pts/5W for RPL and ~7.4pts/10W for MTL.
I end up with ~0.48pts/W for RPL and ~0.74pts/W for MTL.
So MTL is still much more efficient despite RPL needing ~3x longer time to finish that task.

Both of us got something totally different and both of them are flawed.

You get perf/W If you compare them at the same ISO power.
You get W/perf If you compare them at the same ISO performance.
But what we got is some flawed result, because we didn't use ISO power or ISO performance but arbitrarily choose data points to show what we want.

I personally don't have a problem If you compare W/perf instead of perf/W, although I would use both, but at least don't call It as perf/W, because It is not.

 

[Abwx]

You are choosing data points which will show you a result which you want to see.
I can also choose ~2.4pts/5W for RPL and ~7.4pts/10W for MTL.
I end up with ~0.48pts/W for RPL and ~0.74pts/W for MTL.
So MTL is still much more efficient despite RPL needing ~3x longer time to finish that task.

That s my point, by using the same throughput we use the same time, that s the only way to compare intrinsical perf/Watt.

If you increase time you have flawed results, yes, but if you reduce time by increasing frequency you ll have flawed results as well, just in the other direction, hence a same throughput as basis wich mean a same time to perform a same workload.

Curious that you said that a longer time for RPL (compared to MTL) is flawed but didnt catch that a shorter time for the 7840U (compared to MTL) is also flawed...

 

[DavidC1]

I disagree that the P cores are all that matters. It’s entirely possible that Skymont nearly has a 12-14% IPC increase over Gracemont. This gets Skymont pretty close to Zen 3 IPC.

Skymont aims at Golden Cove, not Zen 3. It's triple 3-wide clustered decoder isn't there for a mere 7% increase over Crestmont.

In some aspects Gracemont is already more capable than Golden Cove, such as the dual 32 byte fetch in GMT versus a single 32 byte one in GLC.

One Chinese test of Meteorlake shows that while Intel claimed 4-6% for Crestmont, in his tests Crestmont is 7% faster than the predecessor. I wouldn't be surprised if their goal is another 30% gain over Crestmont just like they did every big core changes.

Crestmont's gain + 30% = Golden Cove. Actually since the 40% is split into 25% Integer and 60% FP, Skymont will probably beat Golden Cove in Integer by 10% or so.

 

[eek2121]

I hope we get some Crestmont only chips like Intel did with Gracemont. The overall performance uplift is minor, but power consumption should be even better. Skylake performance at low wattages is insane!

 

[TESKATLIPOKA]

That s my point, by using the same throughput we use the same time, that s the only way to compare intrinsical perf/Watt.

That's not perf/W but W/perf.

If you increase time you have flawed results, yes, but if you reduce time by increasing frequency you ll have flawed results as well, just in the other direction, hence a same throughput as basis wich mean a same time to perform a same workload.

The flawed result was not only because you increased/decreased the time(performance) needed to finish the job(work, task), but also because the power consumed was not the same. So of course the result was BS.

You can play with both ISO performance and ISO power comparisons, so I find It incomprehensible that you consider only comparing at ISO performance as the correct one, despite It having glaring flaws.

1.flaw) you can set a performance level to one CPU which is not achievable for the other one regardless of how much W It consumes.
2.flaw) If you want to measure efficiency in games, then there you clearly don't measure how fast you finish the game, but how much FPS the CPU can produce per W. So here you are really measuring perf/W and not W/perf.

Curious that you said that a longer time for RPL (compared to MTL) is flawed but didnt catch that a shorter time for the 7840U (compared to MTL) is also flawed...

Just because I didn't comment it doesn't mean I didn't see It or thought It was correct.
At that time I only replied to your first sentence, despite quoting your whole post, that's all.

 

[rtxtwt]

SPECint ST efficiency comparison

SPECint perf graph by same author. IPC regression on MTL is the biggest question mark, tile-base design is to blame?

https://twitter.com/x/status/1741391483578847613

also the normal E core efficiency is good thanks to architechure changes:

https://twitter.com/hjc4869/status/1741687723487056140

There is no such thing like 155H LP P-core.

There IS. MTL has two different type of P core, one of the P core has lower voltage than the original one. Yeah I think it is a painful design.

 

[TESKATLIPOKA]

There IS. MTL has two different type of P core, one of the P core has lower voltage than the original one. Yeah I think it is a painful design.

What? You mean E-cores, right?

 

[mikk]

What? You mean E-cores, right?

He says MTL has two P-cores because of 8VT/6VT Intel 4. The question would be how this works and what core is using the more efficient 6VT. It would make sense the two 1T cores are on 8VT because of the 1T Turbo.

 

[maddie]

So MTL now is believed to have 4 core variations on the die?

 

[FlameTail]

What? You mean E-cores, right?

No.

There are 4 core types.

HV P
LV P
E
LPE

 

[FlameTail]

So how many P cores use 8VT?

 

[Hitman928]

Is the source of those Spec_Int charts just the guy on twitter? The graphs seem somewhat strange to me. Also, what is the source that there are 2 different P cores in MTL? I don't remember hearing about this previously. A 10% IPC reduction in the P cores also doesn't seem correct compared to other tests.

 

[rtxtwt]

Is the source of those Spec_Int charts just the guy on twitter? The graphs seem somewhat strange to me. Also, what is the source that there are 2 different P cores in MTL? I don't remember hearing about this previously. A 10% IPC reduction in the P cores also doesn't seem correct compared to other tests.

What? You mean E-cores, right?

So MTL now is believed to have 4 core variations on the die?

mikk is correct about this. MTL P cores has two different voltage settings. But unfortunately not all reviewers knew this.

https://twitter.com/x/status/1741723232296915143

translate

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

 

[Hitman928]

mikk is correct about this. MTL P cores has two different voltage settings. But unfortunately not all reviewers knew this.

https://twitter.com/x/status/1741723232296915143

translate

This seems to be quoting the same twitter source though? I was asking if there is anyone else making the same claim except this person on twitter. . .

 

[FlameTail]

Are there die shots of MTL CPU tile?

We can have a direct look at the die shot itself to discern if there are two different P cores, because the core with higher voltage will have larger area.

 

[Geddagod]

The curve of the 7840HS is total non sense in respect of the 7840U, that s just made up numbers.

>Sees data he does not like
>Calls data made up numbers
> 🤡

SPECint perf graph by same author. IPC regression on MTL is the biggest question mark, tile-base design is to blame?

Tons of things to potentially blame. Tiles making memory latency/bandwidth/power a "?", spec2017 not liking the RWC's larger L1i cache, I think ringbus frequency got downgraded again? Idk

also the normal E core efficiency is good thanks to architechure changes:

Finally earning the name of "efficient cores" lol

Is the source of those Spec_Int charts just the guy on twitter?

Also, what is the source that there are 2 different P cores in MTL? I don't remember hearing about this previously.

Me neither, but Raichu also alluded to this.

Are there die shots of MTL CPU tile?

We can have a direct look at the die shot itself to discern if there are two different P cores, because the core with higher voltage will have larger area.

Do 6VT and 8VT have area differences? I don't think they do, rather it's channel doping.

Anyway, happy new year!

 

[Hitman928]

>Sees data he does not like
>Calls data made up numbers
> 🤡

Tons of things to potentially blame. Tiles making memory latency/bandwidth/power a "?", spec2017 not liking the RWC's larger L1i cache, I think ringbus frequency got downgraded again? Idk

Finally earning the name of "efficient cores" lol

View attachment 91291


Me neither, but Raichu also alluded to this.

Do 6VT and 8VT have area differences? I don't think they do, rather it's channel doping.

Anyway, happy new year!

Those Spec results are from Geekerwan and they don't agree with the twitter guy's (David Huang) results.

More relevant screenshot here:

 

[tamz_msc]

To add to the discussion - single core performance should be tested with Spec_rate, as it is usually done. However, overall multi-core performance is more realistically captured by Spec_speed.

This is because rate is like Geekbench 5, while speed is like Geekbench 6. The only reason why it isn't tested is because it has a very large memory footprint and it takes too damn long.

 

[Geddagod]

Those Spec results are from Geekerwan and they don't agree with the twitter guy's (David Huang) results.

More relevant screenshot here:

View attachment 91296

View attachment 91295

Both show marginal a marginal IPC decrease in INT tests at higher frequencies. Huang shows a 14% difference between the 155H and 13700H, but he isn't testing IPC afaik, just the scores based on the peak ST frequency, so it's comparing a 13700H at 5GHz vs the 155H at 4.8GHz.
Geekerwan finds a 7% difference at 4.4GHz, and based on how that IPC scaling is going, I wouldn't be surprised if the difference is higher between the two cores at the full boost frequencies of the 155H and 13700H.
I don't find anything too dissimilar or fishy.

 

[Geddagod]

Those Spec results are from Geekerwan and they don't agree with the twitter guy's (David Huang) results.

More relevant screenshot here:

View attachment 91296

View attachment 91295

Oh ma fault, didn't see the edit. Don't know why in these graphs Geekerwan is comparing IPC for RWC at 4.4GHz, and IPC of RPC at 4.7GHz, when his own graphs showed that IPC scaled terribly for RPC between 4.7 and 4.4GHz (<1% increase in perf for 7% increase in frequency)

 

[Abwx]

When a new process is released

>Sees data he does not like
>Calls data made up numbers
> 🤡

You didnt even look at the data, that just a remark out of clulessness; so much for the clownesque argument...

How could the 7840HS score be that lower than the 7840U at say 7-10W, this is exactly the same chip so it should have the same power/perf curve in this part of the graph.

The slope of its power/perf curve is just plain fairy tale since there s a segment where perf increase about linearly with power, whoever doesnt notice such a discrepancy is not qualified to comment the rest...

 

[FlameTail]

twitter guy's (David Huang)

Is he related to Jensen Huang?

 

[Hitman928]

Oh ma fault, didn't see the edit. Don't know why in these graphs Geekerwan is comparing IPC for RWC at 4.4GHz, and IPC of RPC at 4.7GHz, when his own graphs showed that IPC scaled terribly for RPC between 4.7 and 4.4GHz (<1% increase in perf for 7% increase in frequency)

Those are just the default 1t speeds of the respective CPUs. Where is this IPC scaling chart for RPC? At least according to the slides in the new video, RPC scales in IPC slightly better from 3.5 GHz to 4.7 GHz than RWC core scales from 3.6 GHZ to 4.4 GHz.