Discussion Intel Meteor, Arrow, Lunar & Panther Lakes + WCL Discussion Threads

[Tigerick]

Wildcat Lake (WCL) Preliminary Specs

Intel Wildcat Lake (WCL) is upcoming mobile SoC replacing ADL-N. WCL consists of 2 tiles: compute tile and PCD tile. It is true single die consists of CPU, GPU and NPU that is fabbed by 18-A process. Last time I checked, PCD tile is fabbed by TSMC N6 process. They are connected through UCIe, not D2D; a first from Intel. Expecting launching in Q2/Computex 2026. In case people don't remember AlderLake-N, I have created a table below to compare the detail specs of ADL-N and WCL. Just for fun, I am throwing LNL and upcoming Mediatek D9500 SoC.

		Intel Alder Lake - N	Intel Wildcat Lake	Intel Lunar Lake		Mediatek D9500
Launch Date		Q1-2023	Q2-2026 ?	Q3-2024		Q3-2025
Model		Intel N300	?	Core Ultra 7 268V		Dimensity 9500 5G
Dies		2	2	2		1
Node		Intel 7 + ?	Intel 18-A + TSMC N6	TSMC N3B + N6		TSMC N3P
CPU		8 E-cores	2 P-core + 4 LP E-cores	4 P-core + 4 LP E-cores		C1 1+3+4
Threads		8	6	8		8
Max Clock		3.8 GHz	?	5 GHz
L3 Cache		6 MB	?	12 MB
TDP		7 W	Fanless ?	17 W		Fanless
Memory		64-bit LPDDR5-4800	64-bit LPDDR5-6800 ?	128-bit LPDDR5X-8533		64-bit LPDDR5X-10667
Size		16 GB	?	32 GB		24 GB ?
Bandwidth			~ 55 GB/s	136 GB/s		85.6 GB/s
GPU		UHD Graphics		Arc 140V		G1 Ultra
EU / Xe		32 EU	2 Xe	8 Xe		12
Max Clock		1.25 GHz		2 GHz
NPU		NA	18 TOPS	48 TOPS		100 TOPS ?

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.

 

[MS_AT]

Thanks. People still don't seem to realize that HT has drawbacks. One of which is the extra transistor flipping and data movement uses more power and creates more heat. Meaning HT limits clock rates (not a lot, but some). And the efficiency improvements come in-part by no longer having to do the logic checks to keep threads apart on the same core for security reasons (you don't want one thread to be able to access even the cache from another thread).

HT limits clock rates because it is increasing backend utilization, chips draws more power as it is actually doing something so the max clock is limited due thermal/power/current limits. So this extra transistor flipping is generally why you want to have HT in the first place. I mean without HT you will see the chip clocking higher, as the chips are able to achieve very high clocks when they are doing nothing Now, where HT is beneficial depends on the application. If you are able to fill the core with work from single thread, then second thread is useless. In Intel's hybrid case is probably not worth the scheduling difficulties and can save them money on die area they would need to pay TSMC, I find it funny that now ppl start to say HT is useless. It's like Intel's marketing and glued dies, even though they are now using glued dies themselves, eh.

OMG cudimm xmp 9600！ running at gear2 confirmed!View attachment 107469

So much bandwidth wasted on Intel's decision to kill AVX512 in desktop

 

[SiliconFly]

.

 

[9949asd]

HT limits clock rates because it is increasing backend utilization, chips draws more power as it is actually doing something so the max clock is limited due thermal/power/current limits. So this extra transistor flipping is generally why you want to have HT in the first place. I mean without HT you will see the chip clocking higher, as the chips are able to achieve very high clocks when they are doing nothing Now, where HT is beneficial depends on the application. If you are able to fill the core with work from single thread, then second thread is useless. In Intel's hybrid case is probably not worth the scheduling difficulties and can save them money on die area they would need to pay TSMC, I find it funny that now ppl start to say HT is useless. It's like Intel's marketing and glued dies, even though they are now using glued dies themselves, eh.

So much bandwidth wasted on Intel's decision to kill AVX512 in desktop

I guess intel 10nm+avx512=500w, that why they give up avx512. Maybe when intel use tsmc n2 they will add avx512 back.

 

[alcoholbob]

ARL will not show real performance on cpuz now just like zen. On r23 r24 spec will show the correct uplift.

CPU-Z multicore does scale pretty well with R23 multi. If Arrow Lake single core scaling isn't showing on CPU-Z, it may be a sign that the single thread IPC just isn't there or isn't much better than Raptor Lake.

 

[jdubs03]

Spotted this. Just one benchmark, but it shows the U7 258V is fallible..

Intel Core Ultra 7 258V beats AMD Z1 Extreme in DX12 gaming benchmark [Updated]

Lunar Lake comes out on top.

www.tomshardware.com

Packed with 16GB of RAM, the AMD Ryzen Z1 Extreme consistently beats the Intel Core Ultra 7 258V in CPU and average frame rates. The former has 11.3% better performance, delivering 7.5 more frames per second than the latter. And when you put the numbers of AMD’s RDNA 3 Radeon Graphics into the mix, you get a whopping 92.7% more frame rates.

This is an astounding result, especially given that the Intel Core Ultra 7 258V is probably running on a laptop, whereas the AMD Ryzen Z1 Extreme is exclusively available on handheld gaming consoles. While the Intel chip is likely carrying a heavier burden due to the higher resolution (2880x1800 vs. 1920x1080), it does have double the RAM to help it out.

 

[SiliconFly]

.

 

[Hulk]

Yeah may not. But Zen 5 is a flop in uplift with almost no gain in gaming and other consumer workloads that do not use AVX512. Will Arrow Lake actually have a meaningful gain over RPL in gaming and other general consumer workloads even if only 10-15%. I really want to see that plus much lower power for cooler running CPUs and oh stability stability stability which Raptor Lake lacks.

Raptor Lake may even lack stability and still degrade too easily even with corrected microcode and default Intel power profile updates. It was a rushed design and flawed afterall.

No, just normal overclocking instability and failures we've all experienced over the past 30 years of overclocking the brains out of our CPU's.
Difference here is Intel made the default specs too high so they are aggressively overclocked out of the box.

 

[SiliconFly]

.

 

[MS_AT]

I guess intel 10nm+avx512=500w, that why they give up avx512. Maybe when intel use tsmc n2 they will add avx512 back.

They won't as that wouldn't allow them to spam E-cores. Different targets. It's the reason they are pushing AVX10/256. So E-cores will get access to fancy instructions from AVX512 but will be able to do that with double pumped 128b units emulating 256b.

Consumers won't pick number one. They simply pick the best.

There are very few consumers relative to whole consumer market who are able to pick best product. Usually the best marketing wins. That is why Intel is enjoying such large market share in laptops as most of those are sold to enterprises where AMD presence in people's mindset is non-existent. Even for servers it took 4 generations of straight wins from AMD to move the needle.

At their current state, they can't continue to afford to keep burning so much money on multiple designs. Either they have to improve their financials, or go the AMD route.

Multiple designs to introduce artificial market segmentation. This is what basically killed AVX512 market share, and if that would gain foothold in non enterprise usage it would give them additional advantage over ARM competition.

Yeah may not. But Zen 5 is a flop in uplift with almost no gain in gaming and other consumer workloads that do not use AVX512. Will Arrow Lake actually have a meaningful gain over RPL in gaming and other general consumer workloads even if only 10-15%.

Hmm didn't benchmarks for web browsers improve around 15%? Isn't that the most used client software in the world? Or is Arrow Lake 15% improvement in web browsing meaningful and Zen5's a flop.

For servers HT is fine. Not for clients. HT comes with a host of issues like ST performance penalty, efficiency due to lower clocks & power draw, scheduling issues, host of security issues, yuck! Now that HT is out in client, lets pray Intel doesn't bring it back in the future. It just makes the core ugly.

I am waiting till Intel will bring HT back citing efficiency as the reason, eh, this will be fun. You have more security issues thanks to speculative execution than HT, should Intel stop doing that too? At least there will be no misprediction penalty About lower efficiency this is what David Huang measured:

https://blog.hjc.im/wp-content/uploads/2024/08/HX370-MT-SPEC-efficiency-SMTonoff-2048x1281.png

It's simply another tool in the box. Intel now has E-cores so getting rid of HT makes sense from scheduling point of view.

AVX512 is a niche in clients. Pretty much useless for most average consumers.

It's not useless, it's niche in client as is somewhat available in client for only 2 years (since Zen4), before that there were no client capable cores [let's skip Rocket Lake, it was waste of sand and it was replaced by Alder Lake in less than a year] so you have no client software capable of running that as there was no market. Who in right mind would target it... It's gaining foothold slowly now in emulation market, text parsing, making its way to browsers. It's market segmentation that made it niche not it's overall usefulness. Remember it brings new features not only extending width. It's making working with SIMD easier...

 

[cebri1]

Spotted this. Just one benchmark, but it shows the U7 258V is fallible..

Intel Core Ultra 7 258V beats AMD Z1 Extreme in DX12 gaming benchmark [Updated]

Lunar Lake comes out on top.

www.tomshardware.com

Comparing 1080p vs 1800p? Really?

 

[SiliconFly]

.

 

[jdubs03]

Comparing 1080p vs 1800p? Really?

Yeahh, probably jumped the gun a little on that when I saw it. They could've down-scaled the resolution so it'd be comparable.

 

[SiliconFly]

.

 

[jdubs03]

You didn't just jump the gun. You jumped way beyond the gun.

Dare I say, it was cannon fodder?

 

[MS_AT]

Afaik, most of the new SIMD use cases like LLMs now target NPU+GPU instead of CPU with AVX512.

These are the most medial ones due to how popular LLMs have become. Then they can be useful for inferencing, this is Intel marketing angle, as it is less intensive than training and if you need to incorporate the results into something else that is already being done on CPU it kinda makes some sense.

 

[blackangus]

You need to buy a new cudimm xmp9200, hope the price will not be sky high.

Are CUDIMMs usable with any motherboard that support standard DDR5 or do they require memory controller support?
I haven't been keeping up on memory tech.

 

[CouncilorIrissa]

If anyone's looking for bragging rights, I think cinebench uses AVX512. So, if your CPU has it, you'll get better scores.

No.

• Cinebench is a mix of scalar SSE and 256-bit AVX.

CPU-Z multicore does scale pretty well with R23 multi. If Arrow Lake single core scaling isn't showing on CPU-Z, it may be a sign that the single thread IPC just isn't there or isn't much better than Raptor Lake.

CPU-Z is a meme benchmark, the whole dataset fits in the L1 of any modern CPU.
Just like the R23 (why'd you measure general purpose performance in one specific renderer?).

Use actual workloads to measure performance uplift, or at least use benchmarks that represent a mix of workloads (SPEC, GB...)

 

[blackangus]

They have been like that for more than a decade. Only recently have financial problems came up and largely due to Raptor Lake degradation/stability issues.

Intels financial issues are only slight impacted by Raptor Lake. Their major money sink is thier foundry business not being able to perform well.
That has been happening for nearly a decade, and the **** is finally hitting the fan from that with the current economy.
And while its not only the foundry business that is causing them issues it is a large chunk of their problem.
This is why they are thinking about spinning it off.

 

[QuickyDuck]

I would say cpu is not irrelevant for AI workload

 

[Henry swagger]

Yeahh, probably jumped the gun a little on that when I saw it. They could've down-scaled the resolution so it'd be comparable.

Rushing for click bait

 

[Henry swagger]

No.



CPU-Z is a meme benchmark, the whole dataset fits in the L1 of any modern CPU.
Just like the R23 (why'd you measure general purpose performance in one specific renderer?).

Use actual workloads to measure performance uplift, or at least use benchmarks that represent a mix of workloads (SPEC, GB...)

Its better thsn geekbench avx lol

 

[LightningZ71]

They won't as that wouldn't allow them to spam E-cores. Different targets. It's the reason they are pushing AVX10/256. So E-cores will get access to fancy instructions from AVX512 but will be able to do that with double pumped 128b units emulating 256b.

There are very few consumers relative to whole consumer market who are able to pick best product. Usually the best marketing wins. That is why Intel is enjoying such large market share in laptops as most of those are sold to enterprises where AMD presence in people's mindset is non-existent. Even for servers it took 4 generations of straight wins from AMD to move the needle.

Multiple designs to introduce artificial market segmentation. This is what basically killed AVX512 market share, and if that would gain foothold in non enterprise usage it would give them additional advantage over ARM competition.

Hmm didn't benchmarks for web browsers improve around 15%? Isn't that the most used client software in the world? Or is Arrow Lake 15% improvement in web browsing meaningful and Zen5's a flop.

I am waiting till Intel will bring HT back citing efficiency as the reason, eh, this will be fun. You have more security issues thanks to speculative execution than HT, should Intel stop doing that too? At least there will be no misprediction penalty About lower efficiency this is what David Huang measured:

https://blog.hjc.im/wp-content/uploads/2024/08/HX370-MT-SPEC-efficiency-SMTonoff-2048x1281.png

It's simply another tool in the box. Intel now has E-cores so getting rid of HT makes sense from scheduling point of view.

It's not useless, it's niche in client as is somewhat available in client for only 2 years (since Zen4), before that there were no client capable cores [let's skip Rocket Lake, it was waste of sand and it was replaced by Alder Lake in less than a year] so you have no client software capable of running that as there was no market. Who in right mind would target it... It's gaining foothold slowly now in emulation market, text parsing, making its way to browsers. It's market segmentation that made it niche not it's overall usefulness. Remember it brings new features not only extending width. It's making working with SIMD easier...

I guess Ice Lake and Tiger Lake were both just chopped liver. AVX512 was just fine on Tiger Lake h.

 

[jdubs03]

Rushing for click bait

Nah, just glossed over the resolution by mistake. The 258V did have twice the memory, but still not ideal. We’ll get 1 for 1 comparisons soon enough.

 

[511]

I think AVX-512 will die for intel and will come back in AVX 10.2 Form they can have E core with Zen 4 like double pumped avx-512 and P core already have full width AVX-512
Amd follows intel in ISA not the other way around

 

[DavidC1]

I think AVX-512 will die for intel and will come back in AVX 10.2 Form they can have E core with Zen 4 like double pumped avx-512 and P core already have full width AVX-512
Amd follows intel in ISA not the other way around

Maybe in 128-bit support, but that's a stretch. Supporting it itself increases transistor count and power because of the extra registers.

AVX512 is a remnant of the misguided and fat Intel that aimed at continually increasing vector performance of CPUs to limit GPU inroads. There were AVX1024 in plans at one point according to some. It should have went from AVX2 to AVX3, but with 256-bit vector.

The E cores have been closely following ARM cores in development. After all, Atom was shifted to compete with ARM in Smartphones early on and that never stopped.