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]

The hype is cool and such, but we haven't seen reliable 3rd party benchmarks. E-cores cut corners by scaling down the FP side so replacing P-cores is not really likely.

Gracemont shows comparable performance to Skylake and Zen 2. So, I won't be surprised if Skymont will reach the Golden Cove performance, considering significant changes to the architecture.

The key benefit of E-cores is their size. They are nearly 3 times smaller than one P-core. It's much easier to increase the performance just by widening the structures of the small core than optimizing the already big P-core.

 

[FlameTail]

It looks like Skymont's successor will be the next P-core.

Arctic Wolf?

 

[TwistedAndy]

Arctic Wolf?

It depends on Intel.

Skymont is a way more perspective arch than Lion Cove, considering how small it is.

At the same time, it makes sense to keep two design teams competing. Also, it makes sense to keep two lineups with all E- and all P-cores for the server market.

 

[Hulk]

Gracemont shows comparable performance to Skylake and Zen 2. So, I won't be surprised if Skymont will reach the Golden Cove performance, considering significant changes to the architecture.

The key benefit of E-cores is their size. They are nearly 3 times smaller than one P-core. It's much easier to increase the performance just by widening the structures of the small core than optimizing the already big P-core.

Actually they are comparing Skymont to Raptor Cove, which has a little better IPC than Golden Cove due to the larger L2. Crestmont to Skymont looks to be a jump in performance akin to Skylake>Cypress Cove>Golden Cove>Raptor Cove in one jump.

Regarding LP E power in Lunar Lake vs. Meteor Lake. Intel says:

At Crestmont LPE peak performance/power, Skymont will equal that performance at 1/3 the power both ST and MT.

At Crestmont LPE peak performance/power, Skymont using equal power will provide 1.7x more performance ST and 2.9x more performance MT.

When looking at peak vs. peak MT performance for Skymont LPE vs Crestmont LPE (yes, this is 4 core vs. 2), Skymont LPE provides 4x more performance due to better IPC and higher frequency.

If Intel isn't gaslighting us these are huge numbers. We already saw that when workloads could be kept on the LPE in Meteor Lake it was remarkably efficient. The problem was the LPE was so weak not many loads could be contained there. Lunar Lake seems to have fixed this and as a bonus greatly increased efficiency.

Intel looks to have gone over Lunar Lake with the proverbial fine toothed comb. Hopefully actual systems will live up to the hype. One thing that will happen is people will probably say efficiency isn't good when they're testing gaming laptops with power hungry screens and discrete graphics! It's the thin and light configurations I'm interested in seeing benches on. It's going to smash MTL for sure.

 

[coercitiv]

Fun fact, I didn't realize it at the time, but Intel inadvertently gave us the way to calculate a rough estimate for their SMT implementation in terms of area cost.

I'll quote @Saylick 's post because it contains the necessary slides:

But here's the relevant slides:
View attachment 100417
View attachment 100418
View attachment 100419

So according to Intel, performance improves by ~30% when using SMT. At the same time removing SMT logic results in only a 15% perf/area loss, so the SMT core has about 17.5% better perf/area. This means the SMT enabled core is ~10.5% larger.

 

[Hitman928]

Gracemont shows comparable performance to Skylake and Zen 2. So, I won't be surprised if Skymont will reach the Golden Cove performance, considering significant changes to the architecture.

The key benefit of E-cores is their size. They are nearly 3 times smaller than one P-core. It's much easier to increase the performance just by widening the structures of the small core than optimizing the already big P-core.

E-cores are roughly half (just under) the size of P-cores in LNL.

 

[FlameTail]

E-cores are roughly half (just under) the size of P-cores in LNL.

No, it's 1/3rd.

Diagram from @Hulk

 

[Wolverine2349]

It depends on Intel.

Skymont is a way more perspective arch than Lion Cove, considering how small it is.

At the same time, it makes sense to keep two design teams competing. Also, it makes sense to keep two lineups with all E- and all P-cores for the server market.

And for high end desktop market and gaming market as well. I want an all P homogenous non-hybrid arch more than 8 cores for gaming on a single die/node/CCX-CC/ring bus/tile for gaming.

 

[Wolverine2349]

Mediocre in what?
MT looks like It will be better in case you can use every core, single looks similar. For gaming Zen5 with 3D cache.

BTW, which game can use more than 8 cores?

Some parts and situations depending on settings in games Cyber Punk, Spiderman Remastered, The Last of Us Part 1, Cities Skylines if cities get too big.

Also Starfield and Dragons Dogma 2.

Those games can use more than 8 cores. Or am I wrong? I have seen so many conflicting reports that not sure what to believe. I do not have any of those games yet except Cyberpunk and LOU Part 1. I have not progressed far enough into either game to know if they really do use more thna 8 cores or not, but so far only 60% max usage of 7800X3D. But who knows if for a split second intermittently they can expand beyond 8 cores. ALso have not gotten t most NPC intensive parts of game yet so they could use more CPU cores??

 

[Hitman928]

No, it's 1/3rd.
View attachment 101152
Diagram from @Hulk

That’s including L2 for the P-core but not the E-core which isn’t a valid comparison. (Edit: if you take Hulk's E-core cluster estimate and include 1/4 the L2 cache per E-core, it gets to the same ratio that I do).

This is what I got for comparison and including 1/4 of the L2 cache as part of the E-core die estimate. If you leave out the L2 cache for both, the E-core will be even closer to the P-core in size because the P-core has much more L2 cache per core.

Edit: Outside of targeting a lower frequency, Intel's E-cores really aren't small cores anymore.

 

[dullard]

That’s including L2 for the P-core but not the E-core which isn’t a valid comparison.

I agree with that. But I also think that you excluding the L3 cache for the P-core also isn't a valid comparison. That is, unless somehow you want to determine P core performance without the L3 cache when comparing E core to P core performance.

 

[Hitman928]

I agree with that. But I also think that you excluding the L3 cache for the P-core also isn't a valid comparison. That is, unless somehow you want to determine P core performance without the L3 cache when comparing E core to P core performance.

Most comparisons typically go up to L2 cache because all modern cores have at least L2. You do forget that there is still a SLC that the E-cores have access to as well and the E-cores in ARL have access to the L3. You could go all out and just include all cache the core has access to, but then you'd have to include all of the L2 for a single E-core for ST comparisons but less than the full L2 for MT and would also have the SLC vs. L3. Additionally, you'd have this weird dichotomy where the same E-core in LNL and ARL have different area estimates based solely on the shared cache they have access to.

So, I think stopping at L2 is appropriate in terms of comparing core area. Comparing full PPA with/without SLC and L3 requires more nuance.

Edit: I think the cores without L2 or beyond is also an interesting comparison point but if I had to go with just 1, I'd still pick with L2.

 

[ondma]

Mediocre in what?
MT looks like It will be better in case you can use every core, single looks similar. For gaming Zen5 with 3D cache.

BTW, which game can use more than 8 cores?

Everything? Yes, Skymont seems like a great improvement, but I am still going with all the early indications that hyperthreading will be absent, so there goes MT performance. Based on preliminary data from Lunar Lake, it looks like Lion Cove ST performance also does not show a major increase. I am saying this in the perspective that we have had basically 2 refreshes since Alder Lake, one quite good (RL) and one (RL-R) that was a joke. So it is time for a major jump in performance, which I just don't see on the desktop.

Maybe I am setting too high a bar, but after all the hype of: "5 nodes in 5 years", "performance per watt leadership", "AMD in the rear view mirror", and the fact that Zen 5 is coming out first, (possibly with earlier than usual release the 3D chips), I would consider anything but clearly being faster than vanilla Zen 5 to be "mediocre".

 

[H433x0n]

That’s including L2 for the P-core but not the E-core which isn’t a valid comparison. (Edit: if you take Hulk's E-core cluster estimate and include 1/4 the L2 cache per E-core, it gets to the same ratio that I do).

This is what I got for comparison and including 1/4 of the L2 cache as part of the E-core die estimate. If you leave out the L2 cache for both, the E-core will be even closer to the P-core in size because the P-core has much more L2 cache per core.

Edit: Outside of targeting a lower frequency, Intel's E-cores really aren't small cores anymore.

Lion Cove in LNL is the smallest variant of the P-core. I’d imagine Lion Cove in ARL-S is ~15% larger due to the added L2$ and HT structures being present in the core. This would get it closer to that 3:1 ratio.

 

[H433x0n]

Everything? Yes, Skymont seems like a great improvement, but I am still going with all the early indications that hyperthreading will be absent, so there goes MT performance. Based on preliminary data from Lunar Lake, it looks like Lion Cove ST performance also does not show a major increase. I am saying this in the perspective that we have had basically 2 refreshes since Alder Lake, one quite good (RL) and one (RL-R) that was a joke. So it is time for a major jump in performance, which I just don't see on the desktop.

Maybe I am setting too high a bar, but after all the hype of: "5 nodes in 5 years", "performance per watt leadership", "AMD in the rear view mirror", and the fact that Zen 5 is coming out first, (possibly with earlier than usual release the 3D chips), I would consider anything but clearly being faster than vanilla Zen 5 to be "mediocre".

The performance per watt and “5 nodes in 4 years” claims are becoming a reality. They’ve now got the intel 4/3 node out to market. They’re halfway there and the timelines haven’t slipped yet.

As for performance of the architectures, I would wait for 3rd party reviews before declaring anything mediocre.

 

[Wolverine2349]

o.

The performance per watt and “5 nodes in 4 years” claims are becoming a reality. They’ve now got the intel 4/3 node out to market. They’re halfway there and the timelines haven’t slipped yet.

As for performance of the architectures, I would wait for 3rd party reviews before declaring anything mediocre.

I hope performance is really strong with Lion Cove. Is a 20% IPC improvement across all work loads realistic? COuld Intle be sandbagging with 14% IPC uplift claims. Given clock speed regression, I would like to see at least 20% IPC uplift with minor clock speed regression at worst.

 

[H433x0n]

I hope performance is really strong with Lion Cove. Is a 20% IPC improvement across all work loads realistic? COuld Intle be sandbagging with 14% IPC uplift claims. Given clock speed regression, I would like to see at least 20% IPC uplift with minor clock speed regression at worst.

I think that if you expect anything greater than 14% you’re looking for disappointment.

 

[Wolverine2349]

I think that if you expect anything greater than 14% you’re looking for disappointment.

IPC or overall performance of Lion Cove cores?

Will there be any clock speed regression or maybe not as was once thought?

 

[Hitman928]

Lion Cove in LNL is the smallest variant of the P-core. I’d imagine Lion Cove in ARL-S is ~15% larger due to the added L2$ and HT structures being present in the core. This would get it closer to that 3:1 ratio.

I'm not convinced that ARL will have HT circuitry, but even if we do add 15% to the P-cores, you're still looking at a 1:2.5 ratio rather than 1:3.

 

[dullard]

Maybe I am setting too high a bar, but after all the hype of: "5 nodes in 5 years", "performance per watt leadership", "AMD in the rear view mirror", and the fact that Zen 5 is coming out first, (possibly with earlier than usual release the 3D chips), I would consider anything but clearly being faster than vanilla Zen 5 to be "mediocre".

I think you need to go back and think about those claims in context of what was actually being spoken about. When Intel claimed 5 nodes in 4 years, Intel did not claim that every chip along the way would be a leadership chip. It was always spoken about in terms of the future. This graph has been shared many times here, but it needs to be stickied on the top.

It isn't until 18A that Intel is claiming performance/watt leadership and even then they listed it as the same (≈) or a bit better (+). It is 14A when Intel claims outright leadership, not 20A with Arrow Lake. And even then Intel didn't list 14A with two plus signs of being far superior. If you end up disappointed, it is only because you misheard the claims in the first place.

Arrow Lake won't be a bad chip. But, it will still be made on approximately* the same node size as Zen 5. Arrow Lake can't be clearly faster in all metrics based on that alone. *I say approximately since there will be multiple tiles on multiple nodes, and there isn't a direct comparison.

Based on preliminary data from Lunar Lake, it looks like Lion Cove ST performance also does not show a major increase.

I don't understand why you'd think a 15 W ultralight laptop CPU should be taken as the basis of performance for a ~125 W desktop CPU.

 

[ondma]

I think you need to go back and think about those claims in context of what was actually being spoken about. When Intel claimed 5 nodes in 4 years, Intel did not claim that every chip along the way would be a leadership chip. It was always spoken about in terms of the future. This graph has been shared many times here, but it needs to be stickied on the top.
View attachment 101186
It isn't until 18A that Intel is claiming performance/watt leadership and even then they listed it as the same (≈) or a bit better (+). It is 14A when Intel claims outright leadership, not 20A with Arrow Lake. And even then Intel didn't list 14A with two plus signs of being far superior. If you end up disappointed, it is only because you misheard the claims in the first place.

Arrow Lake won't be a bad chip. But, it will still be made on approximately* the same node size as Zen 5. Arrow Lake can't be clearly faster in all metrics based on that alone. *I say approximately since there will be multiple tiles on multiple nodes, and there isn't a direct comparison.


I don't understand why you'd think a 15 W ultralight laptop CPU should be taken as the basis of performance for a ~125 W desktop CPU.

Arrow Lake needs to be a lot better than "not bad". It needs to be a home run, especially since it is coming out after Zen 5. ARL might have been a great chip if it had come out instead of RL-R, ahead of Zen 5. But coming out several months after Zen 5 and maybe even the X3D chips, puts it in a very bad position. And it even doesn't use the fancy new intel nodes for the most performant chips, but TSMC. Honestly, Intel seems to still be in a state of complete disarray, especially on desktop. Maybe Lunar Lake will be great for low power laptops, but I dont see any advance like that on the desktop.

 

[dullard]

Arrow Lake needs to be a lot better than "not bad". It needs to be a home run, especially since it is coming out after Zen 5.

I see no reason that Intel needs Arrow Lake to be a home run. Intel did okay financially with lackluster nodes compared to TSMC in the last couple of years. Profit was drastically down, but they were still profitable (and despite lower profits were still double the profit in 2023 of the much larger AMD--by market cap). Intel will also do financially okay with nodes that are on roughly par with TSMC (and even when using TSMC nodes).

The real test to me is Panther Lake. It powered on last week for the first time. Intel has put just about all of its eggs into the 18A basket. If Panther Lake doesn't cut it, then Intel may be in a world of hurt. Quite a few important companies have signed onto 18A.

 

[Hitman928]

I see no reason that Intel needs Arrow Lake to be a home run. Intel did okay financially with lackluster nodes compared to TSMC in the last couple of years. Profit was drastically down, but they were still profitable (and despite lower profits were still double the profit in 2023 of the much larger AMD--by market cap). Intel will also do financially okay with nodes that are on roughly par with TSMC (and even when using TSMC nodes).

The real test to me is Panther Lake. It powered on last week for the first time. Intel has put just about all of its eggs into the 18A basket. If Panther Lake doesn't cut it, then Intel may be in a world of hurt. Quite a few important companies have signed onto 18A.

Intel was not profitable for 3/4 quarters last year and barely broke even with a profitable 4th quarter.

Intel annual operating income for 2023 was $0.093B

Last I saw, they weren't projecting a profit for 2024 either, but that may change with the AI craziness happening. That's just operating income too, if you look at their actual cash flow, they are spending billions more annually for the last couple of years than they are bringing in, trying to get their fab business built up.

Intel annual free cash flow for 2023 was $-14.279B
Intel annual free cash flow for 2022 was $-9.617B

 

[dullard]

Intel was not profitable for 3/4 quarters last year and barely broke even with a profitable 4th quarter.



Last I saw, they weren't projecting a profit for 2024 either, but that may change with the AI craziness happening. That's just operating income too, if you look at their actual cash flow, they are spending billions more annually for the last couple of years than they are bringing in, trying to get their fab business built up.

Intel net income for 2023: $1.689B for a $129B company.
AMD net income for 2023: $0.854B for a $258B company.
Neither did particularly well last year--exactly what I stated.

Yes, cash flow is negative as they are dumping money into 18A (and beyond). Also as I stated. If 18A doesn't pan out, Intel is in for a world of hurt. But, investing for the future is a good sign for a company.

 

[Hitman928]

Intel net income for 2023: $1.689B for a $129B company.
AMD net income for 2023: $0.854B for a $258B company.
Neither did particularly well last year--exactly what I stated.

Yes, cash flow is negative as they are dumping money into 18A (and beyond). Also as I stated. If 18A doesn't pan out, Intel is in for a world of hurt. But, investing for the future is a good sign for a company.

So you think we should ignore operating expenses? Is labor free all of a sudden?