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.

 

[Meteor Late]

Power efficiency still pretty meh even with 3nm and all these "efficiency" cores. Multi threaded performance also not that amazing, basically ties zen 5 average.

People seem to be severely over estimating e-cores. The only thing they seem good for is optimizing the die area for Cinebench r24.

E cores are less efficient than P cores at the upper part of the curve. Yes less efficient, not more or even equal, there is data out there that shows this, so on Desktop, they are worse for MT than P cores, another thing is that of course their much lower area allows Intel to stuff more of them.

 

[igor_kavinski]

Intel really needed to keep hyperthreading. It's easy to see why Intel is behind without it.

AGREED

Their biggest lie: "We axed HT for single threaded performance"

WHAT single thread performance????

You are literally losing in almost 99% of games!!!!

Like, you don't wanna hear my shrieking screaming voice right now, Intel!

Lion Cove's expanded structures probably only work to the max with HT enabled!

 

[desrever]

E cores are less efficient than P cores at the upper part of the curve. Yes less efficient, not more or even equal, there is data out there that shows this, so on Desktop, they are worse for MT than P cores, another thing is that of course their much lower area allows Intel to stuff more of them.

My point is, people here act like E-core should replace P-core but they can't even clock to 4.6ghz without completely tanking efficiency.

Even if they can significantly improve IPC for the e-core next gen, if they are still clocked at <5 ghz good luck replacing P-core as a design.

 

[Hans Gruber]

AGREED

Their biggest lie: "We axed HT for single threaded performance"

WHAT single thread performance????

You are literally losing in almost 99% of games!!!!

Like, you don't wanna hear my shrieking screaming voice right now, Intel!

Lion Cove's expanded structures probably only work to the max with HT enabled!

Let's go back to 9th generation Intel. They dropped Hyperthreading because of security issues. Without hyperthreading no security issues. Hyperthreading returned for the 10th generation Intel chips through 14th generation. Hyperthreading will probably return for Panther Lake (18A).

 

[Joe NYC]

Hopefully someone tests gaming with e-cores disabled. That should shed some light on whether the windows scheduler is the problem

Isn't the APO thing supposed to know what to do with the E-Cores? From Hardware Nexus, it did not seem to make much difference, even on the games that are supposed to support it.

 

[Thunder 57]

Let's go back to 9th generation Intel. They dropped Hyperthreading because of security issues. Without hyperthreading no security issues. Hyperthreading returned for the 10th generation Intel chips through 14th generation. Hyperthreading will probably return for Panther Lake (18A).

Segmentation issues. The 9900k still had HT.

 

[cannedlake240]

My point is, people here act like E-core should replace P-core but they can't even clock to 4.6ghz without completely tanking efficiency.

Even if they can significantly improve IPC for the e-core next gen, if they are still clocked at <5 ghz good luck replacing P-core as a design.

Well that's why it's an E core. It's not designed for high clock speed. E cores won't replace P cores, they'll have to redesign them for a proper replacement

 

[Ranulf]

Hey where did you guys get all the bad reviews ?
285k is 1st on userbenchmark with 38% perf over 14900ks.


View attachment 110219

If you buy a 285k next week for the low low price of $629.95... you're gonna party like its the tops of 1466. Cinebench or AD, take your pick. Okay, okay...like its 2021 AL, aka Alder Lake.

Top parts Intel. Top...parts.

 

[igor_kavinski]

285K stock DDR5-6400 Phoronix results:

Spoiler: I'll be the JUDGE

Speedometer 3: Ah waht?? Pathetic fail
Jetstream 2: Same. FAIL
OSPRAY: Fail
Embree: Fail
IndigoBench Bedroom: Fail
Intel Open Image Denoise: Fail
Appleseed Emily: Fail
V-ray: Fail
LuxCore Benchmark: Fail
ACES DEGEMM: Fail
miniBUDE: Fail
GROMACS: FAIL
NAMD: FAIL
Xmrig: Fail
Clickhouse: Fail
DuckDB Clickbench: Fail
simdjson: Fail
Numpy: Fail
Cryptsetup: Fail (Decisive win only in Serpent-XTS and fails elsewhere)
SVT-AV1: Fail (despite being great in one test)
x265: Fail (because it is miserable in one test)
Kvazaar: Fail
uvg266: Fail
LibRAW: Fail (loses to 9700X!)
Liquid-DSP: Fail
Ngspice: Fail
srsRAN: Fail (despite being good in one test)
TensorFlow: Fail
OpenVINO: Fail
Whisper.cpp: Fail
Tested games: FAIL

WASM collisiondetection: Decisive Win
WASM imageconvulate: Decisive Win
Godot compilation: Decisive Win
LLVM compilation: Decisive Win
Mesa compilation: Decisive Win
CoreMark: Decisive Win
QuantLib: Decisive Win
BRL-CAD: Decisive Win
libxsmm: Decisive Win
GPAW: Decisive Win
Xcompact3D: Decisive Win
SPECFEM3D: Decisive Win
nginx: Decisive Win
Apache: Decisive Win
Cpuminer: VERY Decisive Win (despite slighly slower in one test)
Apache IoTDB: Decisive Win
PGSQL: Decisive Win (despite losing in Read Only but most real world DB workloads are rarely read only)
CockroachDB: Decisive Win
PyBench: Decisive Win
PyPerformance: Decisive Win
PHPBench: Decisive Win
SVT-VP9: Decisive Win
WebP: Decisive Win
C-ray: Decisive Win

Linux kernel compilation: Great and 2nd only to 9950X
FFmpeg compilation: Great and barely faster than 9950X
PHP compilation: Great and 2nd only to 9950X
Gem5 compilation: Great and 2nd only to 9950X
7-zip compression: Great and 2nd only to 9950X
Blender BMW: Great and 2nd only to 9950X
Blender Junkshop: Great and 2nd only to 9950X
LuxCoreRender DLSC: Great and 2nd only to 9950X
LuxCore Orange Juice: Great and 2nd only to 9950X
Appleseed Disney: Great and 2nd only to 9950X
IndigoBench Supercar: Great and 2nd only to 9950X
DuckDB TPC-H: Great and 2nd only to 9950X
ASTC Encoder: Great and 2nd only to 9950X
Memcached: Great and 2nd only to 9950X
OpenFOAM: Much Great and 2nd only to 7950X3D (quite a feat!)
Pennant: Great and 2nd only to the X3Ds

Blender Fishycat: Just OK
Blender Pabellon Barcelona: Just OK

LuxCoreRender Rainbow: Barely wins
LAMMPS: Barely wins against 9950X

OpenRadioss: Overall great with 1 loss and two wins against 9950X
NAS Parallel Benchmarks: Overall great with 1 loss and two great wins against competitors

RocksDB: Serviceable (Stellar in random read but loses in Read while write, the latter being more important)

 

[Jan Olšan]

AGREED

Their biggest lie: "We axed HT for single threaded performance"

WHAT single thread performance????

You are literally losing in almost 99% of games!!!!

Like, you don't wanna hear my shrieking screaming voice right now, Intel!

Lion Cove's expanded structures probably only work to the max with HT enabled!

Obviously the answer is, it would have been even worse with HT (or at least in Lunar).


Seriously, lack of HT is not Arrow Lake's problem. All the guys saying it is loosing in benchmarks due to that: Do you realise that HT only covered third of the cores? If HT could add 15 % performance on P-Cores, that becomes +5% and likely less due to non-linear scaling in the whole picture. Hardly relevant.

Consider that the core was designed with even hybrid-er hybrid designs, like 8+32 (20 % of cores affected) in mind. For Intel's hybrid design, it is pretty much true or close enough that you can offset the removal of SMT with adding more little/efficient cores, because the area saved can likely buy some.
Add to that that your scheduling decisions are simplified from P-Core × E-Core × do you populate HT threads on P-Cores or not to just P-Core × E-Core. The scheduling still isn't a fully solved issue, so perhaps this helps enough to offset some of that lost 5% MT performance.

It's different for AMD because they have SMT on all cores, so removal would cost them much more performance.

 

[Meteor Late]

E cores are good for stuffing many of them AND clocking them low so that efficiency is similar or slightly higher than P cores. We are talking 3-3.5GHz kind of low at maximum. Basically, they are a laptop type of core, efficient and cheap MT, coupled with at least two P core for smooth day to day workloads, ideally 4 at least.
Put them in Desktop on a high power limit CPU and they will not be that good.
E cores are a really good idea, just in the wrong segment and pushed way out of their sweet spot in order to chase AMD's high core count parts.

 

[511]

285K stock DDR5-6400 Phoronix results:

Spoiler: I'll be the JUDGE

Speedometer 3: Ah waht?? Pathetic fail
Jetstream 2: Same. FAIL
OSPRAY: Fail
Embree: Fail
IndigoBench Bedroom: Fail
Intel Open Image Denoise: Fail
Appleseed Emily: Fail
V-ray: Fail
LuxCore Benchmark: Fail
ACES DEGEMM: Fail
miniBUDE: Fail
GROMACS: FAIL
NAMD: FAIL
Xmrig: Fail
Clickhouse: Fail
DuckDB Clickbench: Fail
simdjson: Fail
Numpy: Fail
Cryptsetup: Fail (Decisive win only in Serpent-XTS and fails elsewhere)
SVT-AV1: Fail (despite being great in one test)
x265: Fail (because it is miserable in one test)
Kvazaar: Fail
uvg266: Fail
LibRAW: Fail (loses to 9700X!)
Liquid-DSP: Fail
Ngspice: Fail
srsRAN: Fail (despite being good in one test)
TensorFlow: Fail
OpenVINO: Fail
Whisper.cpp: Fail
Tested games: FAIL

WASM collisiondetection: Decisive Win
WASM imageconvulate: Decisive Win
Godot compilation: Decisive Win
LLVM compilation: Decisive Win
Mesa compilation: Decisive Win
CoreMark: Decisive Win
QuantLib: Decisive Win
BRL-CAD: Decisive Win
libxsmm: Decisive Win
GPAW: Decisive Win
Xcompact3D: Decisive Win
SPECFEM3D: Decisive Win
nginx: Decisive Win
Apache: Decisive Win
Cpuminer: VERY Decisive Win (despite slighly slower in one test)
Apache IoTDB: Decisive Win
PGSQL: Decisive Win (despite losing in Read Only but most real world DB workloads are rarely read only)
CockroachDB: Decisive Win
PyBench: Decisive Win
PyPerformance: Decisive Win
PHPBench: Decisive Win
SVT-VP9: Decisive Win
WebP: Decisive Win
C-ray: Decisive Win

Linux kernel compilation: Great and 2nd only to 9950X
FFmpeg compilation: Great and barely faster than 9950X
PHP compilation: Great and 2nd only to 9950X
Gem5 compilation: Great and 2nd only to 9950X
7-zip compression: Great and 2nd only to 9950X
Blender BMW: Great and 2nd only to 9950X
Blender Junkshop: Great and 2nd only to 9950X
LuxCoreRender DLSC: Great and 2nd only to 9950X
LuxCore Orange Juice: Great and 2nd only to 9950X
Appleseed Disney: Great and 2nd only to 9950X
IndigoBench Supercar: Great and 2nd only to 9950X
DuckDB TPC-H: Great and 2nd only to 9950X
ASTC Encoder: Great and 2nd only to 9950X
Memcached: Great and 2nd only to 9950X
OpenFOAM: Much Great and 2nd only to 7950X3D (quite a feat!)
Pennant: Great and 2nd only to the X3Ds

Blender Fishycat: Just OK
Blender Pabellon Barcelona: Just OK

LuxCoreRender Rainbow: Barely wins
LAMMPS: Barely wins against 9950X

OpenRadioss: Overall great with 1 loss and two wins against 9950X
NAS Parallel Benchmarks: Overall great with 1 loss and two great wins against competitors

RocksDB: Serviceable (Stellar in random read but loses in Read while write, the latter being more important)

Intel loosing at their own software

 

[511]

intel should Just discount Raptor lake CPUs and sell them it is made on depriciated node that is the only way to make money and Launch Bartlett Lake ASAP with AVX-512

 

[gdansk]

E cores are good for stuffing many of them AND clocking them low so that efficiency is similar or slightly higher than P cores. We are talking 3-3.5GHz kind of low at maximum. Basically, they are a laptop type of core, efficient and cheap MT, coupled with at least two P core for smooth day to day workloads, ideally 4 at least.
Put them in Desktop on a high power limit CPU and they will not be that good.

E is perhaps a misnomer. Perhaps D for Dense.

Skymont is likely a partial inspiration of their next P and E cores. But don't expect it to be small. The advantage is the front-end which seems most likely to punch through the current x64 instruction level parallelism limitations and is scalable for both P and E designs.

 

[Hitman928]

Isn't the APO thing supposed to know what to do with the E-Cores? From Hardware Nexus, it did not seem to make much difference, even on the games that are supposed to support it.

Do you mean Gamer's Nexus? They tested APO on and off but there was no difference for the 285k.

 

[GodisanAtheist]

 

[majord]

What happened to Thread director? Didnt it solve scheduling issue years ago, and shouldn't it be even better now?

 

[igor_kavinski]

Seriously, lack of HT is not Arrow Lake's problem.

Check out Phoronix's review. HT would've given them a lot of MT wins against 9950X and even X3D CPUs. It could've been a very decent workstation CPU for development and other professional tasks.

 

[igor_kavinski]

What happened to Thread director? Didnt it solve scheduling issue years ago, and shouldn't it be even better now?

The Thread Director misses the HT threads. It's so gloomy now. The Director's heart aches like nothing you know.

 

[Abwx]

Intel really needed to keep hyperthreading. It's easy to see why Intel is behind without it.

With SMT they would had gained the equivalent of 2P assuming a paltry 25% SMT gain, that s not negligible, eventualy they ll get back to this approach in their next designs as the bigger a core the more the SMT relevance.

 

[DavidC1]

Check out Phoronix's review. HT would've given them a lot of MT wins against 9950X and even X3D CPUs. It could've been a very decent workstation CPU for development and other professional tasks.

HT didn't happen because they were executing badly on Lion Cove.

 

[igor_kavinski]

HT didn't happen because they were executing badly on Lion Cove.

Hope they realize their mistake and release the Refresh with HT enabled next year.

 

[DavidC1]

Hope they realize their mistake and release the Refresh with HT enabled next year.

A mistake is not doing things on purpose and realizing later that decision was wrong.

This is incompetence.

The advantage is the front-end which seems most likely to punch through the current x64 instruction level parallelism limitations and is scalable for both P and E designs.

There's more details than just the front-end.

 

[GTracing]

Isn't the APO thing supposed to know what to do with the E-Cores? From Hardware Nexus, it did not seem to make much difference, even on the games that are supposed to support it.

Yeah, it's supposed to, but certain games/benchmarks are abnormally bad, and a bad scheduler seems to be the most likely cause.

 

[DavidC1]

Yeah, it's supposed to, but certain games/benchmarks are abnormally bad, and a bad scheduler seems to be the most likely cause.

And this:

they are still clocked at <5 ghz good luck replacing P-core as a design.

Which is what they should be doing. Cut the clocks to ~5GHz, which will allow reducing of pipeline stages, tightening memory and cache latencies, and end up with a better core that's fit for desktop, server, and mobile.

Why do you think they had to back down ring clocks drastically? Cause they couldn't clock it that high without turning into Raptorlake.