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

Tigerick · Aug 22, 2022

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.)

adroc_thurston · Mar 19, 2024

FlameTail said:
GEMM?

GEneral Matrix Multiply.
Matrix math!
The very foundation of machine learning.

H433x0n said:
Was this test at JEDEC-4800?

ARL is 6400.

H433x0n said:
Expectations are pretty low but if IPC bump is <15% then they deserve to get clobbered.

They are lol.

itsmydamnation · Mar 19, 2024

FlameTail said:
GEMM?

matrix multiply

GEMM units are better at GEMM then SIMD because SIMD writes back to register after every op , GEMM do it after whatever depth the GEMM unit is.

igor_kavinski · Mar 19, 2024

adroc_thurston said:
The very foundation of machine learning.

Or you know everything visual like images/videos.

igor_kavinski · Mar 19, 2024

itsmydamnation said:
because SIMD writes back to register after every op ,

What's the drawback if it doesn't do that after every op?

igor_kavinski · Mar 19, 2024

adroc_thurston said:
ARL is 6400.

ARL may be decent if paired with CL28 DDR5-6400 RAM.

Goop_reformed · Mar 20, 2024

Geddagod said:
Lmao Xino said ARL will be the same perf as RPL or maybe a singe digits level improvement (prob referring to ST)

This can't be real right?

S'renne · Mar 20, 2024

Goop_reformed said:
This can't be real right?

If the focus is on efficiency it could be, seeing how bad the inefficiency Raptor Lake is at the moment. Could be better in MT than Zen 5 but worse in ST and edges out in efficiency kind of deal

adroc_thurston · Mar 20, 2024

Goop_reformed said:
This can't be real right?

Lmao.

S'renne said:
If the focus is on efficiency it could be

Intrinsically every core focuses on efficiency.

S'renne said:
seeing how bad the inefficiency Raptor Lake is at the moment

It's ok unless you consider KS parts real.
Raptor-HX is genuinely alright.

S'renne said:
Could be better in MT than Zen 5 but worse in ST and edges out in efficiency kind of deal

It doesn't.
Skymont is pretty good but it's no Zen5.

S'renne · Mar 20, 2024

adroc_thurston said:
Lmao.

Intrinsically every core focuses on efficiency.

It's ok unless you consider KS parts real.
Raptor-HX is genuinely alright.

It doesn't.
Skymont is pretty good but it's no Zen5.

Just saying the "efficiency emphasis" idea out loud only because of the rumour of LNL having 1.5x MT GB(? Forgot which software) score than MTL or something like that

itsmydamnation · Mar 20, 2024

igor_kavinski said:
What's the drawback if it doesn't do that after every op?

If you want to do gemm nothing it's why the exsist, if you want to do something other then gemm well your doing your math in the wrong unit.

Gemm units are specialised for a specific workload , avx/simd units are general for any work that can be batched together.

Ghostsonplanets · Mar 20, 2024

adroc_thurston said:
Skymont is pretty good but it's no Zen5.

It's interesting that while there a lot of rumors about ARC not being a good jump over GLC, the rumors around Skymont are pretty positive. It's almost as if the Atom team has been able to deliver a consistent cadence of performance improvements while the Core team takes their sweet time.

Tremont to Gracemont was a substantial jump, Crestmont was a nice update over Gracemont and Skymont is rumored to be around GLC performance, which is really good for what are meant to be area focused cores. One might wonder if the Atom team won't surpass the Core team someday if they keep the impressive cadence.

igor_kavinski · Mar 20, 2024

Ghostsonplanets said:
Crestmont was a nice update over Gracemont

Have you seen any benchmarks comparing these two? Would you like to share?

Ghostsonplanets · Mar 20, 2024

igor_kavinski said:
Have you seen any benchmarks comparing these two? Would you like to share?

I haven't, sadly. Hopefully Chips and Cheese will do one down the road.

I'm just going by the Intel claim of 7 - 8% IPC upgrade of Crestmont over Gracemont.

moinmoin · Mar 20, 2024

Ghostsonplanets said:
It's almost as if the Atom team has been able to deliver a consistent cadence of performance improvements while the Core team takes their sweet time.

Technology wise the Atom team's work indeed has been more interesting to follow for quite some time now, like for a decade by now?

Goop_reformed · Mar 20, 2024

adroc_thurston said:
Lmao.

What do you mean lmao. Talk to me like I am a 5 year old or a golden retriever.

Geddagod · Mar 20, 2024

Goop_reformed said:
What do you mean lmao. Talk to me like I am a 5 year old or a golden retriever.

That's wild lmaoooo

dullard · Mar 20, 2024

adroc_thurston said:
?
They do bf/fp8|16 math just fine.

Most support FP16/BF16 just fine.

They can support different data types. But, they often do not. This link below is an Intel graphic, so it is obvious that it is geared to put Intel in the best light, but you can see that not all NPUs ran FP16 at the time of the image creation. If anyone can find AMD's list of supported NPU data types, then I'll link that as well (my Google skills are failing me, the latest information I found is that the list of supported data types has not yet been released).

https://cdn.wccftech.com/wp-content/uploads/2023/12/2023-12-14_19-47-23-1456x819.png

naukkis · Mar 20, 2024

adroc_thurston said:
?
They do bf/fp8|16 math just fine.

Most support FP16/BF16 just fine.

Speculation was about replacing FPU with NPU. FPU doesn't usually even support FP16 math, it's single precision is 32 bit and double precision 64 bit(or more). OK, FPU SIMD today do support also FP16 but that's just outlier for something like AI where NPU is aimed at. So no, you can't replace FPU with NPU as they are designed to totally different kind of precision.

Markfw · Mar 20, 2024

dullard said:
They can support different data types. But, they often do not. This link below is an Intel graphic, so it is obvious that it is geared to put Intel in the best light, but you can see that not all NPUs ran FP16 at the time of the image creation. If anyone can find AMD's list of supported NPU data types, then I'll link that as well (my Google skills are failing me, the latest information I found is that the list of supported data types has not yet been released).

https://cdn.wccftech.com/wp-content/uploads/2023/12/2023-12-14_19-47-23-1456x819.png

forbidden on that link

dullard · Mar 20, 2024

Markfw said:
forbidden on that link

Before I post the image, am I forbidden to post it, or are you forbidden to view it?

Markfw · Mar 20, 2024

dullard said:
Before I post the image, am I forbidden to post it, or are you forbidden to view it?

When I click on the link it says forbidden,.

dullard · Mar 20, 2024

naukkis said:
Speculation was about replacing FPU with NPU. FPU doesn't usually even support FP16 math, it's single precision is 32 bit and double precision 64 bit(or more). OK, FPU SIMD today do support also FP16 but that's just outlier for something like AI where NPU is aimed at. So no, you can't replace FPU with NPU as they are designed to totally different kind of precision.

The ability to use different data types is theoretically possible with NPUs. But, the main issue is memory, bandwidth, and power used. If you only need 4 bits (which AI often only needs 4 bits or 8 bits), then using something set up for 512 bits is quite a waste. Using 512 bits when your application needs 4 bits will require 128x more memory, will have to move 128x more data around, and will have to process 128x more of that data, using much more power. All while only being able to use much smaller AI models due to those limits. So, it isn't really efficient to use something set up for 512 bits with 4 bits

The reverse is true too. If you have an NPU optimized and designed for say, 4-bit math, and need 16-bit data, then you need to transfer that data around in 4 chunks which takes more time. Then you have memory to store only 1/4th the data. It can work, but just won't be as performant as you want.

dullard · Mar 20, 2024

Markfw said:
When I click on the link it says forbidden,.

No problem, here is the image. Again, if someone has a list of AMD supported NPU data types, I'll include the official list since I prefer my posts to be neutral in sources.

From here: https://wccftech.com/intel-meteor-l...t-gen-cpu-cores-arc-gpu-npu-ai-pc-revolution/

Edit: here is the list of AMD supported data types: https://www.xilinx.com/products/technology/ai-engine.html

adroc_thurston · Mar 20, 2024

S'renne said:
Just saying the "efficiency emphasis" idea out loud only because of the rumour of LNL having 1.5x MT GB(? Forgot which software) score than MTL or something like that

That's compared to (non-existent) MTL-U which is one of the worst parts Intel has ever made.
Not a real achievement per se.

dullard said:
Again, if someone has a list of AMD supported NPU data types, I'll include the official list since I prefer my posts to be neutral in sources.

this really says nothing about the hardware, just DirectML support.
AIE-ML in PHX supports FP16/BF16 just fine.
see the docs.

adroc_thurston · Mar 20, 2024

dullard said:
Again, if someone has a list of AMD supported NPU data types, I'll include the official list since I prefer my posts to be neutral in sources.

this really says nothing about the hardware, just DirectML support.
AIE2 in PHX supports FP16/BF16 just fine.
see the docs.

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

Senior member

Attachments

Platinum Member

Platinum Member

Lifer

Lifer

Lifer

Member

Member

Platinum Member

Member

Platinum Member

Senior member

Lifer

Senior member

Diamond Member

Member

Golden Member

Elite Member

Senior member

Moderator Emeritus, Elite Member

Elite Member

Moderator Emeritus, Elite Member

Elite Member

Elite Member

Platinum Member

Platinum Member