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.)

 

[sgs_x86]

Very true. But I'd like to remind that 285K is still 3.5% faster than 9950X.

I know 3.5% gain isn't much and might even fall within margin of error, but it's very important to note that for the first time in many years, Intel has come out with a product that feels truly competitive and is pretty much neck-to-neck with competition. Meaning, definitely not lagging significantly anymore. Also, equal footing is apt too.

Alder and Raptor Lakes were competitive too. But consumed a lot of power. Let's hope Arrow is the first step towards correcting that negative.

 

[SiliconFly]

Alder and Raptor Lakes were competitive too. But consumed a lot of power. Let's hope Arrow is the first step towards correcting that negative.

True. They were performant but very power inefficient compared to competition.

 

[AcrosTinus]

I still don't know what GB6 really shows in multi core, everything seems have a ceiling at around 20K points, is it a scaling problem of the bench ?

hopefully this translates well into general performance.

 

[Abwx]

I still don't know what GB6 really shows in multi core, everything seems have a ceiling at around 20K points, is it a scaling problem of the bench ?

hopefully this translates well into general performance.

43118 pts at 250W in the Intel favourable CB R23, wich mean that at 200W it will be outmatched by the 9950X even in this bench, and at its rated 250W it will be no match in Blender, Corona, Vray and likely X264, eventually not even in 7 Zip where the 9950X did few progress if any relatively to the 7950X, so what will be left..?.
Ordering news versions of CB and GB.?.

 

[AcrosTinus]

43118 pts at 250W in the Intel favourable CB R23, wich mean that at 200W it will be outmatched by the 9950X even in this bench, and at its rated 250W it will be no match in Blender, Corona, Vray and likely X264, eventually not even in 7 Zip where the 9950X did few progress if any relatively to the 7950X, so what will be left..?.
Ordering news versions of CB and GB.?.

I mean even the 7950x had some wins against the 14900K if things became really MT heavy, hence my remark that the GB6 scores are too clustered in MT to anything.

Again, this is not Gracemont, the SIMD is significantly beefed up, meaning it is still an unknown whether the 9950x really wins or not.
It might defeat the 9950X in MT which leaves it with nothing but AVX512 and 99% of the apps I use don't use it and if I need to tinker with AVX512, I have a Xeon Server for that.

 

[dullard]

I still don't know what GB6 really shows in multi core, everything seems have a ceiling at around 20K points, is it a scaling problem of the bench ?

hopefully this translates well into general performance.

Geekbench 6 made a significant change to how it handles multi-threaded scores. Look up Amdahl's law: https://en.wikipedia.org/wiki/Amdahl's_law

Older versions of GB (and many other software benchmarks) simply make multiple (N) copies of the same task and run them all at once. So, for a 8-core benchmark, it would run 8 single instances of the same thing all at once. In general, until there is some limitation (like memory bandwidth), there is no real limit to scores. If you run 2 of the same thing in parallel, it is roughly 2x what you'd get when you run 1 of that thing. If you run 8 instances of the same task in parallel, the final result is roughly 8x the single instance score. This method of measuring multi-threaded performance doesn't follow Amdahl's law since it never asks the software to try to break up a task into multiple threads. All instances are just one thread and at no point does the thread have to wait for other threads' results.

But GB6 changes it up*. It now runs 1 task and asks it to use N number of threads. Not a lot of software can actually utilize a large number of threads at the same time. Thus, once you reach a certain number of threads, you get the point of diminishing returns in many types of software. At some point, thread E is waiting for results from lets say thread B or G before it can proceed. Thus, thread E sits idle until it gets the necessary data from the other threads. Amdahl's law kicks in and all results end up hitting a ceiling where all threads, no matter how many cores you have, are waiting for just a single thread to complete before they can resume.

Now the question comes up: are you mostly running N copies of the same software doing the same task all at once? For example, do you have 16 copies of Photoshop running all processing the same image with the same filter all at the same time (distributed computing is a good example where you are doing this)? Or are you mostly running a single software intensely--hoping to process one image as quickly as possible (probably most users)?


*See bottom of page 7: https://www.geekbench.com/doc/geekbench6-benchmark-internals.pdf

 

[Abwx]

I mean even the 7950x had some wins against the 14900K if things became really MT heavy, hence my remark that the GB6 scores are too clustered in MT to anything.

Not only some but the majority of the wins, in Computerbase previous MT tests that comprised 13 benches even the 14900KS at 330W had trouble matching the 7950X, at 253W it was 5-6% behind on average, that s the starting point for the 9950X wich is 9% faster than the 7950X, so that s 15-16% required from ARL to be on par with the AMD opponent.

Again, this is not Gracemont, the SIMD is significantly beefed up, meaning it is still an unknown whether the 9950x really wins or not.
It might defeat the 9950X in MT which leaves it with nothing but AVX512 and 99% of the apps I use don't use it and if I need to tinker with AVX512, I have a Xeon Server for that.

The 43318 score in CB at 250W say it all, we can even extract SKT perf from this number, wich is 4% better IPC in CB R23 than Zen 3, a rare win because in CB R15 and a lot of other benches it could well have a slightly lower IPC.

Beside in almost all benches where the 7950X was dominant in respect of RPL there s no AVX512, only on rarities like y-Cruncher and a few other apps that are never used in benches.

Now when it comes to ST ARL could have a slight win in CB, wich is irrelevant for renderers that are used in MT anyway, and test at Compterbase show the 9950X in good light for lowly threaded tests like WebXPRT, an Intel bench btw, while TPU confirm this behaviour with Speedometer, two benches where ARL early numbers dont match RPL.

 

[AcrosTinus]

Geekbench 6 made a significant change to how it handles multi-threaded scores. Look up Amdahl's law: https://en.wikipedia.org/wiki/Amdahl's_law

Older versions of GB (and many other software benchmarks) simply make multiple copies of the same task and run them all at once. So, for a 8-core benchmark, it would run 8 single instances of the same thing all at once. In general, until there is some limitation (like memory bandwidth), there is no real limit to scores. If you run 2 of the same thing in parallel, it is roughly 2x what you'd get when you run 1 of that thing. If you run 8 instances of the same task in parallel, the final result is roughly 8x the single instance score. This method of measuring multi-threaded performance doesn't follow Amdahl's law since it never asks the software to try to break up a task into multiple threads. All instances are just one thread and at no point does the thread have to wait for other threads' results.

But GB6 changes it up. It now runs 1 task and asks it to use N number threads. Not a lot of software can actually utilize a large number of threads at the same time. Thus, once you reach a certain number of threads, you get the point of diminishing returns in many types of software. At some point, thread E is waiting for results from lets say thread B or G before it can proceed. Thus, thread E sits idle until it can proceed. Amdahl's law kicks in and all results end up hitting a ceiling where all threads, no matter how many cores you have, are waiting for just a single thread to complete before they can resume.

Now the question comes up: are you mostly running N copies of the same software doing the same task all at once? For example, do you have 16 copies of Photoshop running all processing the same image with the same filter all at the same time? Or are you mostly running a single software intensely--hoping to process one image as quickly as possible?

Thank you, that makes sense.

 

[mikk]

Sinners rejoice, the E-core voltage rail deliverance is finally upon us!

https://twitter.com/x/status/1826072744120955299

Only one VCCCORE says Jaykihn.

There’s still only one vcccore input rail, it’s just that it’s further split via DLVR.

P, E, and ring are still a singular input rail.

I think the much more exciting implication is per-core voltage for the P cores (per-module voltage for E cores).

https://twitter.com/x/status/1826083875828936974

 

[Jan Olšan]

Nope. Check the videocardzGB6 numbers (all stock). 9950X is 3335 & 285K is 3450. That puts 285K around 3.5% ahead of 9950X.

I think that is low score for 9950X, ComputerBase has 3427 at DDR5-5600 (not the best memory). Seems it could be a tie instead. We don't know if this is a typical 285K score, a low one or a high one.

Also, there are some results for Geekbench 5. The P-core in Intel Core 285K matches the one in Apple M4:

View attachment 105879

That's not a fully fair comparison because of the Geekbench Windows tax, but it is pretty interesting.

2513 ST / 28214 MT would be a win in multithread, but a loss in single thread in our testing, we got 2586 / 25890 on 9950X (DDR5-6000, but different methods, testing on W10, so who knows what would 285K score under the conditions).

 

[DavidC1]

Older versions of GB (and many other software benchmarks) simply make multiple (N) copies of the same task and run them all at once. So, for a 8-core benchmark, it would run 8 single instances of the same thing all at once. In general, until there is some limitation (like memory bandwidth), there is no real limit to scores.

Sure there is. Look at server CPU tests and Geekbench 5 and older ones scale horribly. You basically ignore Geekbench MT results beyond 16 threads or something. So now it scales even worse and since it's basically an arbitrary numbering system the semi-useless feature is now even more useless.

Geekbench 4 had a sort function, separate memory, cryptography, Int and FP section for both ST and MT. Geekbench 6 skews it with almost nonexistent AVX512 instruction, can't be sorted anymore, and screws up MT.

Surely the "per clock" is totally irrelevant when talking about IPC (Instructions Per Cycle).

Per clock = "IPC" because it's the questionable but widely accepted in AT forums terminology. The term confuses all kinds of people because they think "instruction" and automatically turn into monkeys thinking "oh noes ARM and x86 instructions not same!" nonsense.

@SiliconFly

Ok. It's time for me prepare a "I told you so" list of all the people who said this ARL generation will only get a lower single digit % performance uplift in ST. Whereas, it's already showing 11.7% ST gains against previous gen equivalent 14900K and this isn't final silicon yet!

There are many people in the list, with whom shall I start?

It is BEYOND silly to argue about 2-3% differences when Intel stated 14% improvement and you can extrapolate using 5.7GHz and 6GHz top ST clocks.

On top of that you guys are using Geekbench, a user submitted benchmark that can easily have 50% differences in the extreme case.

Wait for reviews I said. I also said 9% based on 90% clock scaling and 14% gains per clock with 5.7GHz vs 6GHz. 11% is margin of error on a well setup review, nevermind Geekbench.

 

[dullard]

Sure there is. Look at server CPU tests and Geekbench 5 and older ones scale horribly. You basically ignore Geekbench MT results beyond 16 threads or something. So now it scales even worse and since it's basically an arbitrary numbering system the semi-useless feature is now even more useless.

Of course there is a limit eventually. I even mentioned one (memory bandwidth). So your "sure there is" sounds a bit out of place. Maybe you misread my sentence. It was of the form "Until X occurs, Y will happen." To respond with "sure there is" X just seems odd.

All benchmarks are in some form arbitrary. What image is being processed, which game is being played, which data set is being crunched? But to call it useless just puts you in a bad light. It now runs software the way most users run software: one program at a time trying to go as fast as the CPU will let it. Unless you are regularly doing multiple separate single threaded tasks simultaneously, that is far more applicable way to measure CPU performance. Yes, there are use cases where you run many simultaneous single thread tasks at the same time (many server loads for example). But if you are using Geekbench to measure server speeds, I would say your thought process is what was most useless.

Below is what Geekbench measures. It has nothing to do with server-type tasks. Instead it is focused on typical day-to-day "geek" tasks.
"using workloads that include data compression, image processing, and machine learning. Performance on these workloads is important for a wide variety of applications including web browsers, image editors, and developer tools."

Tasks include: file compression, navigation path optimization, popular website browsing, PDF map rendering, image databasing, code compiling, database text processing, compression, machine learning, object detection, background blur, image editing, ray tracing, and 3D structure motion.

Any use case primarily outside those tasks should not use Geekbench as a benchmark.

 

[Henry swagger]

https://twitter.com/x/status/1826472529629729098

impressive 😃

 

[poke01]

https://twitter.com/x/status/1826472529629729098

impressive 😃

The clocks are wrong. GB doesnt always report the right clocks

 

[jdubs03]

https://twitter.com/x/status/1826472529629729098

impressive 😃

That MT score is lower than the other past results. I’d say 20000 would be okayish at best. Even the reported 19000 is barely an uptick. Looks meh. The 285K is the standout from what’s out there so far.

 

[Abwx]

The clocks are wrong. GB doesnt always report the right clocks

The GB linked page state 5497 MHz.

BIOSTAR Group Z890 VALKYRIE - Geekbench

Benchmark results for a BIOSTAR Group Z890 VALKYRIE with an Intel Core Ultra 7 265KF processor.

browser.geekbench.com

 

[poke01]

The GB linked page state 5497 MHz.

BIOSTAR Group Z890 VALKYRIE - Geekbench

Benchmark results for a BIOSTAR Group Z890 VALKYRIE with an Intel Core Ultra 7 265KF processor.

browser.geekbench.com

I’m talking about the listed frequencies in the .gb5 link

 

[igor_kavinski]

https://www.reddit.com/user/BookinCookie/

BookinCookie
• 12h ago

That’s not the nature of Intel’s problem. It’s not about the low-level technical decisions about specific products. The real problem is that Intel’s management has made poor strategic bets, provided inconsistent funding, and disbanded talented teams which could have helped Intel maintain its leadership. Essentially, management is making Intel an unattractive place for talented engineers to work.
And to your other point: yes, their management is that bad. For example their killing of Royal, one of their most innovative projects, is a nearly incomprehensible decision. Both in terms of throwing away their best plan for CPU leadership and in terms of losing many of their best engineers.

From Saylick's shared post:

https://www.reddit.com/r/intel/comments/1eyny65/any_other_intel_employees_here_how_are_yall

 

[jdubs03]

From Saylick's shared post:

https://www.reddit.com/r/intel/comments/1eyny65/any_other_intel_employees_here_how_are_yall

Yea that's not a great post there. It seems like Bookin isn't an Intel employee though? So its 2nd hand grapevine chatter, it seems. But, they do make a good point in one of their posts about the AADG:

https://www.reddit.com/r/intel/comments/1eyny65/comment/ljgt2sr

^That is such a self-own it sounds. Royal Core seemed like it could've been revolutionary in design. Such a shame that their management could mess that up just because of fawning over AI.

The others posts doesn't sound good either. there is definitely a strong whiff of low morale. And the last thing they can afford is their most talented people bouncing. They need to stem the tide or it could be a vicious downward spiral.

 

[AcrosTinus]

Yea that's not a great post there. It seems like Bookin isn't an Intel employee though? So its 2nd hand grapevine chatter, it seems. But, they do make a good point in one of their posts about the AADG:

https://www.reddit.com/r/intel/comments/1eyny65/comment/ljgt2sr

^That is such a self-own it sounds. Royal Core seemed like it could've been revolutionary in design. Such a shame that their management could mess that up just because of fawning over AI.

The others posts doesn't sound good either. there is definitely a strong whiff of low morale. And the last thing they can afford is their most talented people bouncing. They need to stem the tide or it could be a vicious downward spiral.

Royal core is nothing but a rumor, nothing concrete ever leaked like architecture details, just it is going to be fast, trust me bro.

Intel will survive, they have to otherwise I have lost 10K.

 

[Ghostsonplanets]

Yea that's not a great post there. It seems like Bookin isn't an Intel employee though? So its 2nd hand grapevine chatter, it seems. But, they do make a good point in one of their posts about the AADG:

https://www.reddit.com/r/intel/comments/1eyny65/comment/ljgt2sr

^That is such a self-own it sounds. Royal Core seemed like it could've been revolutionary in design. Such a shame that their management could mess that up just because of fawning over AI.

The others posts doesn't sound good either. there is definitely a strong whiff of low morale. And the last thing they can afford is their most talented people bouncing. They need to stem the tide or it could be a vicious downward spiral.

This match with what Exist reported some weeks ago. Unless this person is just regurgitating Exist info.

Hopefully Artic Wolf and Coyote Cove aren't some half-hearted successors to SKT and LNC.

 

[AcrosTinus]

This match with what Exist reported some weeks ago. Unless this person is just regurgitating Exist info.

Hopefully Artic Wolf and Coyote Cove aren't some half-hearted successors to SKT and LNC.

It in fact does not exist. In the lab where I work we have several codenames for products that start, fail and the lesson learned are merged with the next one. A R&D exercise without a concrete instance is not a product and therefore does not exist. I cannot get sad about a potential something that never had any details but a youtuber or leaker telling me something great was in the works but canceled, now smear Intel, they were to able to launch this rumored product they cannot execute...

To me that logic is nonsense!

I look at launched products and evaluate for my needs:
Sandybridge -> epic
Haswell -> epic
broadwell -> alright
Skylake -> epic
Skylake again -> alright
Skaylake again again -> tired
Sunny Cove -> are we back
Cypress Cove -> WTF...
Willow Cove (Tiger Lake) -> epic
Golden Cove (Alder Lake) -> epic
Redwood Cove -> ???
Golden Cove again -> tired

 

[SiliconFly]

CapFrameX posted a new 285K GB5 benckleaks result:

https://twitter.com/x/status/1826186324946149408

It’s consistent with the previous GB6 285K benchmark leak.

He measures ARL-S PPC gain at a whopping 18%.

This puts ARL clearly ahead of competition (even after accounting for the mythical future update that competition may or may not receive and/or will or will not give it a extreme performance bump).

 

[Henry swagger]

CapFrameX posted a new 285K GB5 benckleaks result:

https://twitter.com/x/status/1826186324946149408

It’s consistent with the previous GB6 285K benchmark leak.

He measures ARL-S PPC gain at a whopping 18%.

This puts ARL clearly ahead of competition (even after accounting for the mythical future update that competition may or may not receive and/or will or will not give it a extreme performance bump).

I want cinebench r23 now.. plus the new imc and ddr5 10k will be mega

 

[SiliconFly]

I want cinebench r23 now.. plus the new imc and ddr5 10k will be mega

Well, from what I gather from some of the previous posts, it looks like Lunar Lake’s LNC has 14% IPC uplift & ARL’s LNC has 18% IPC uplift. It’s more than expected.