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.

 

[Thunder 57]

Platform Tops

https://twitter.com/x/status/1887856347498320330

Does anyone care about TOPS though? I thought the concensus was that it was a waste of die space though could be used for something more valuable. As far as the NPU goes that is.

 

[GTracing]

Platform Tops

https://twitter.com/x/status/1887856347498320330

How meaningful is platform TOPS as a metric? Genuinely asking. I don't know much about this stuff.

Can models easily be run on the CPU, NPU, and GPU at the same time? Will a 10 TOPS CPU run AI software as fast as a 10 TOPS NPU or 10 TOPS GPU? Or does it depend on the software? Would running a model on the CPU use all the CPU horsepower and make the system slow and unusable until the software finishes?

 

[511]

Platform Tops

https://twitter.com/x/status/1887856347498320330

From this info the clock speed for iGPU is around 2.5 GHz and there is a core count increase of 50% for the halo and the TOPs per Xe Cores are same so
2.5*4096(Int 8)*12 = 122 TOPS

 

[jpiniero]

Does anyone care about TOPS though?

Wall Street does

 

[511]

How meaningful is platform TOPS as a metric? Genuinely asking. I don't know much about this stuff.

TOPS is even worse than FP32 Flops for metric cause at least it's FP32 and is fixed Size while companies don't even list the TOPS they use it can be Int8/BF16/FP16/Int4/FP4 size.

Nvidia uses TOPS with an edge case called Sparsity they will just take the TOPS multiply by 2 due to the edge case and report it as TOPS for example the 980 TOPs on Nvidia RTX 5070 is FP4 with Sparsity so if we convert it into what Intel has reported it in terms of Int8 we would be dividing it by 4. 2 from going from FP4(4 bit) Sparsity to non Sparsity and 2 by going from FP4 to Int8 (8 bit) that would translate to approx 245 TOPS of Int8 or roughly double the throughput of the Panther Lake iGPU.

TOPS = ops/clk*clk frequency* Cores
This is the mathematical part of conversion

Can models easily be run on the CPU, NPU, and GPU at the same time? Will a 10 TOPS CPU run AI software as fast as a 10 TOPS NPU or 10 TOPS GPU? Or does it depend on the software? Would running a model on the CPU use all the CPU horsepower and make the system slow and unusable until the software finishes?

It all comes down to software also if they are not Memory Bandwidth and Size bound it would be performing the same but that's an ideal scenario an iGPU may be faster than an RTX 4090 if the model size is too big that it spills out of VRAM and becomes Slow.
For AI You need few things
> Memory large enough to fit the model.
> Memory Bandwidth to not slow the model
> Compute ofc measured in TOPS in this case BF16 is used for training smaller size like FP4 is limited to inference.
> Software Support( Hence the CUDA Moat cause AMDs MI300X is more powerful but lacking in SW)

Hope that helps

 

[511]

Does anyone care about TOPS though? I thought the concensus was that it was a waste of die space though could be used for something more valuable. As far as the NPU goes that is.

Btw Intel has cut the NPU Size in Half from the rumors

 

[dullard]

Yes, E cores are quite competent at SPECINT2017 at low clock speeds compared to P cores, but just to point out the obvious....

• Real applications are not SPECINT

That test was SPECint. But, there have been other tests of P core vs. E core on other applications. The results so far have always the same: the performance curves cross each other at some point. Meaning one core is better at low power and the other core is better at high power.

Here are some admittedly dated ones, but they were easy to find and are real applications and not just SPECint:

Alder Lake’s Power Efficiency – A Complicated Picture

Reviews across the internet show Alder Lake getting very competitive performance with very high power consumption.

chipsandcheese.com

but my point is for power constrained devices P cores are not that useful outside of maybe like 4P Cores the additional 2P cores will steal too much from power budget.

If your revised statement is that 6 P cores is too much, then lucky for you the 235H exists with fewer P cores and costs 31% less.

 

[dullard]

Btw Intel has cut the NPU Size in Half from the rumors

Going from TSMC N6 to Intel 18A, they better be able to cut it in half.

 

[coercitiv]

Btw Intel has cut the NPU Size in Half from the rumors

Perfectly balanced, as all things should be.

 

[511]

Going from TSMC N6 to Intel 18A, they better be able to cut it in half.

Not from N6 to 18A I meant N3B to 18A

 

[dullard]

Not from N6 to 18A I meant N3B to 18A

I was thinking of the NPU on Meteor Lake and Arrow Lake which uses TSMC N6. I assume you were talking about Lunar Lake?

 

[511]

I was thinking of the NPU on Meteor Lake and Arrow Lake which uses TSMC N6. I assume you were talking about Lunar Lake?

Yes they have made a significant changes compared to LNL to minimize the area for NPU while still meeting Co-Pilot requirements

 

[OneEng2]

That test was SPECint. But, there have been other tests of P core vs. E core on other applications. The results so far have always the same: the performance curves cross each other at some point. Meaning one core is better at low power and the other core is better at high power.

Here are some admittedly dated ones, but they were easy to find and are real applications and not just SPECint:
View attachment 116511
View attachment 116512

Alder Lake’s Power Efficiency – A Complicated Picture

Reviews across the internet show Alder Lake getting very competitive performance with very high power consumption.

chipsandcheese.com

If your revised statement is that 6 P cores is too much, then lucky for you the 235H exists with fewer P cores and costs 31% less.

Thanks. This is for older P and E cores and not LLK/ARL, but I suspect the story would still be the same.

 

[511]

Thanks. This is for older P and E cores and not LLK/ARL, but I suspect the story would still be the same.

The architecture has massive gaps between GLC and GMT vs SKT and LNC so there is a massive uArch difference the curve has shifted in favour of SKT more vs LNL

 

[cannedlake240]

Thanks. This is for older P and E cores and not LLK/ARL, but I suspect the story would still be the same.

Nope, since then e cores got >20% ppc in integer, even more in fp and higher clocks while P cores only got 10% ppc

 

[LightningZ71]

There is certainly room in the mobile stack for a 4xlion plus 16xskt.

 

[fastandfurious6]

Intel Core Ultra 9 275HX fails to outperform Core i9-14900HX in Cinebench R23 single-core test

A Cinebench R23 score of Intel's Arrow Lake-based Core Ultra 9 275HX has emerged online. While it fails to best its Raptor Lake counterpart in single-core performance, it offers a double-digit increment in the multicore test.

www.notebookcheck.net

🫣🫣🫣

 

[fastandfurious6]

isn't it strange both camps still keep review embargos etc despite all products details announced already month+ ago?

when embargos lifted?

 

[AcrosTinus]

Intel Core Ultra 9 275HX fails to outperform Core i9-14900HX in Cinebench R23 single-core test

A Cinebench R23 score of Intel's Arrow Lake-based Core Ultra 9 275HX has emerged online. While it fails to best its Raptor Lake counterpart in single-core performance, it offers a double-digit increment in the multicore test.

www.notebookcheck.net

🫣🫣🫣

Raptor is and was fast, so fast in fact it burned out like a star.
Only 18A or 14A based chips with a lot more logic can really surpass it, Zen5 is not much faster than RaptorCove as well.

 

[511]

Raptor is and was fast, so fast in fact it burned out like a star.
Only 18A or 14A based chips with a lot more logic can really surpass it, Zen5 is not much faster than RaptorCove as well.

It's embarrassing that a Intel 7 product is making ST Performance of Zen5/ARL not meh

 

[511]

I am waiting for Igor to leak NVL Performance Projection like they did with arrow lake.

Since the core count is 16P+32E I bet the 32E alone would have greater Multi than ARL-S and 2X over ARL is possible 100K R23?

I hope they don't use DLVR on Desktop there is no need for it on Desktop.

Any guesstimate for ST Performance?

 

[Thunder 57]

I am waiting for Igor to leak NVL Performance Projection like they did with arrow lake.

Since the core count is 16P+32E I bet the 32E alone would have greater Multi than ARL-S and 2X over ARL is possible 100K R23?

I hope they don't use DLVR on Desktop there is no need for it on Desktop

At what kind of TDP do you think they could pull that off? And enough memory bandwitdth to feed the cores? Maybe they have a halo SKU for that but I suspect Intel will stick to 8P cores for the most part. Not necssarily a bad thing, look at Nvidia ATM.

 

[511]

At what kind of TDP do you think they could pull that off? And enough memory bandwitdth to feed the cores? Maybe they have a halo SKU for that but I suspect Intel will stick to 8P cores for the most part. Not necssarily a bad thing, look at Nvidia ATM.

Around 300-350W which kind of is needed to feed the P cores if it were 48 E cores 250W would have been enough considering 144E SRF Cores are fine around 250W as for bandwidth I would guess Triple channel memory ought to be needed

 

[igor_kavinski]

Around 300-350W

I think they might choose default TDP of 270 to 280W but unlimited with 420mm AIO may go up to 400W.

 

[511]

I think they might choose default TDP of 270 to 280W but unlimited with 420mm AIO may go up to 400W.

yeah but i worry about the P cores