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.)

 

[Henry swagger]

Intel highlights Lunar Lake, Xeon 6 and Gaudi 3 at Hot Chips 2024 - VideoCardz.com

Intel Demonstrates AI Architectural Expertise at Hot Chips 2024 Intel Xeon 6 SoC and Lunar Lake processors, alongside Intel Gaudi 3 AI accelerators and optical compute interconnect, headline technical presentations. What’s New: Demonstrating the depth and breadth of its technologies at Hot Chips...

videocardz.com

Granite rapids uses redwood cove plus which is 8 wide.. wonder how much ipc it gets vs lion cove

 

[DavidC1]

The cross core latency is interesting. The E core latency is similar to P core latency on MTL, and the P core latency is half of MTL. They say the low latency is one factor that enables the E cores to be used for multi-threaded workloads.

Getting FHD screen instead of OLED higher resolution has another long-term advantage. The system will run cooler and the battery will last longer. Modern laptops get way too hot, which is a killer for batteries. My Yoga never gets more than warm, and it's completely passive. The fan doesn't bother me but a fan means it needs vents, meaning it gets dusty. HWInfo is saying after 7-8 years the degradation is only 10-15% for the battery which is excellent.

Intel highlights Lunar Lake, Xeon 6 and Gaudi 3 at Hot Chips 2024 - VideoCardz.com

Intel Demonstrates AI Architectural Expertise at Hot Chips 2024 Intel Xeon 6 SoC and Lunar Lake processors, alongside Intel Gaudi 3 AI accelerators and optical compute interconnect, headline technical presentations. What’s New: Demonstrating the depth and breadth of its technologies at Hot Chips...

videocardz.com

Granite rapids uses redwood cove plus which is 8 wide.. wonder how much ipc it gets vs lion cove

Can always refer back to what Pat said.
"10-plus percent on the core..."

So that alone will make it more competitive with Turin, since it's not client RWC.

Interesting they are saying Intel 4 for the IO die on GNR-D. Probably a mistake.

 

[poke01]

So memory latency is still bad

 

[dullard]

Interesting they are saying Intel 4 for the IO die on GNR-D. Probably a mistake.

All the slides say Intel 4 for the IO die. https://wccftech.com/intel-granite-...o-perf-2x-ethernet-throughput-up-to-42-cores/

Note: press the arrow buttons on the link to see more slides.

 

[DavidC1]

All the slides say Intel 4 for the IO die. https://wccftech.com/intel-granite-...o-perf-2x-ethernet-throughput-up-to-42-cores/

Note: press the arrow buttons on the link to see more slides.

Which is weird. And you can't completely trust them, because they claimed 192EU for Meteorlake, but according to some that was never the case. Or what about calling Sapphire Rapids die a Sierra Forest die?

Intel 4 is supposed to not have IO capability. Intel 3 does.

So memory latency is still bad

Typical of laptops though. This is part of why Desktop version is 5% or so faster, not some fabric latency issue.

 

[poke01]

Typical of laptops though.

I thought it would improve in Lunar lake considering the on-package memory. Apple’s M4 latency is around 90ns which is significantly better than Lunar’s.

Is there any reason why?

 

[DavidC1]

I thought it would improve in Lunar lake considering the on-package memory. Apple’s M4 latency is around 90ns which is significantly better than Lunar’s.

Is there any reason why?

Moving the memory on package itself won't lower latency. Otherwise, the lowest memory latency would be on Smartphones and Tablets. The lowest is on desktops, not Apple M4. On package memory is mostly about saving space. Actually Lunarlake talking about PHY power reduction shows even power savings won't be much unless some re-engineering is done.

While signal integrity will improve allowing it to potentially clock it higher, latency terms it's limited by DRAM architecture itself. The physical latency portion is an insignificant amount. HBM is even closer but latency is no better. It has loads of bandwidth though.

 

[majord]

5-6W with screen off isn't particularly efficient. If WiFi 7 takes that much power I don't think it's worth it. Based on those numbers alone it's no advancement at all from my Kabylake on the power front. But Meteorlake system shows nearly double so we'll see.

It's a bit puzzling why they would demo such poor total consumption for either machine.. I know it's supposed to be a comparison video, but anyone with any sort of off-top-of-the head idea on what are 'good' numbers isn't going to be impressed.


Jaykhin's Package power numbers look pretty good I guess, nothing surprising though considering Meteor lake wasn't exactly the pinnacle.

 

[poke01]

The lowest is on desktops, not Apple M4

I know lowest is on desktop. I was more referring to the closest Lunar implementation outside Intel which is in tablets/laptops. It’s a big difference regarding the memory latency on M4 and Lunar. ~90ns vs ~128ns is quite big. Both LPDRR5.

Raptor mobile is around 110ns. Looks to be a design choice by Intel.

 

[ikjadoon]

I know lowest is on desktop. I was more referring to the closest Lunar implementation outside Intel which is in tablets/laptops. It’s a big difference regarding the memory latency on M4 and Lunar. ~90ns vs ~128ns is quite big. Both LPDRR5.

Raptor mobile is around 110ns. Looks to be a design choice by Intel.

I've noticed some DRAM latency tests can be rather noisy, e.g., perhaps Intel's LNL test is not comparable with Geekerwan's M4 test?

Even within one test, it can be quite variable on the device / test. And maybe even run?

In AIDA64's DRAM latency test,

this i7-1250U + LPDDR5-5200 laptop hit 89 ns
this i7-1250U + LPDDR5-5200 laptop hit 146 ns (+64%)

No idea why. Timings appear similar enough, though one is downclocked temporarily, and the timings tightened commensurately. Unrelatedly, Qualcomm's Oryon DRAM latency is erroneously low (5ns to 10ns) in the same AIDA64 test.

What's a good real-world test of latency impact? Maybe better than synthetics.

//

You bring up a good point. With LNL, OEMs won't have a choice of LPDDR5X SKUs: everybody gets the same DRAM (I presume Intel won't be offering multiple latency SKUs per capacity option).

//

Somewhat unrelatedly, on-package LPDDR5 does use a different package size than soldered LPDDR5. I know little about this. Perhaps on-package DRAM SKUs get worse bins due to their lower volume?

Square-ish shape vs rectangle:

 

[quarph]

I know lowest is on desktop. I was more referring to the closest Lunar implementation outside Intel which is in tablets/laptops. It’s a big difference regarding the memory latency on M4 and Lunar. ~90ns vs ~128ns is quite big. Both LPDRR5.

Raptor mobile is around 110ns. Looks to be a design choice by Intel.

It is probably not an apples to apples comparison. Intel's result here is with 4KB page size; it includes address translation penalty spilling the TLB coverage.

 

[511]

The cross core latency is interesting. The E core latency is similar to P core latency on MTL, and the P core latency is half of MTL. They say the low latency is one factor that enables the E cores to be used for multi-threaded workloads.

Getting FHD screen instead of OLED higher resolution has another long-term advantage. The system will run cooler and the battery will last longer. Modern laptops get way too hot, which is a killer for batteries. My Yoga never gets more than warm, and it's completely passive. The fan doesn't bother me but a fan means it needs vents, meaning it gets dusty. HWInfo is saying after 7-8 years the degradation is only 10-15% for the battery which is excellent.

Can always refer back to what Pat said.
"10-plus percent on the core..."

So that alone will make it more competitive with Turin, since it's not client RWC.

Interesting they are saying Intel 4 for the IO die on GNR-D. Probably a mistake.

According to One Raichu they mixed LNC front end with RWC+

https://twitter.com/x/status/1828258538583732299

 

[poke01]

Somewhat unrelatedly, on-package LPDDR5 does use a different package size than soldered LPDDR5. I know little about this. Perhaps on-package DRAM SKUs get worse bins due to their lower volume?

It can't be due to lower volume as M3 Max has a square package. I think Lunar and M3 just have that shape due to space constraints.

 

[jdubs03]

According to One Raichu they mixed LNC front end with RWC+

https://twitter.com/x/status/1828258538583732299

It’s hidden to us mortals. 🥺

 

[FlameTail]

View attachment 106293

It can't be due to lower volume as M3 Max has a square package. I think Lunar and M3 just have that shape due to space constraints.

M3 Max uses 128b LPDDR modules.

Base M3 and Lunar Lake use 64b LPDDR modules.

So that explains why the shape is different.

 

[511]

It’s hidden to us mortals. 🥺

 

[Doug S]

I've noticed some DRAM latency tests can be rather noisy, e.g., perhaps Intel's LNL test is not comparable with Geekerwan's M4 test?

You can't compare different latency tests, they aren't measuring the same thing. Gotta use the same software and testing methodology for both.

Not all software that purports to measure latency is accurately measuring it. Some older software may "defeat" prefetch with varying levels of success (for instance Apple Silicon's prefetch algorithm appears to be able to detect backward and larger step strides others do not) or does not understand the difference between x86's 4K pages and Apple's 16K pages as @quarph pointed out. Or even if you've accounted for that different CPUs have different TLB sizes and levels, but if you want to measure memory latency you don't want to add TLB (especially L2 TLB) latency to your figures.

So basically I'm not buying any comparison of Intel vs Apple latency measured in different ways. Heck even using the SAME software it can be tricky and you might think you've done it right then someone who knows some of the deep dark secrets of a platform like Hans DeVries or @name99 will point out something you weren't taking into account. Basically, there isn't a "true" result in measured in nanoseconds for a given platform any more than there's a "true" result of IPC measured in instructions per cycle. But you definitely can't claim two measurements done using different software by different people have any validity as a comparison whatsoever.

 

[DavidC1]

@511 It would be sad indeed if the 8-wide means it's from the uop cache like Bionic_Squash suggests. Truly the marketing team's garbage has infested both companies.

 

[511]

It would be sad indeed if the 8-wide means it's from the uop cache like Bionic_Squash suggests. Truly the marketing team's garbage has infested both companies.

Yeah indeed i have mixed feelings regarding the 35% layoff from sales/marketing it is necessary but it's sad to see people getting laid off anyway we need someone to bench it but if it had IPC gains they would clearly mention it even if it was single digit like crestmont only time will tell

 

[DavidC1]

Yeah indeed i have mixed feelings regarding the 35% layoff from sales/marketing it is necessary but it's sad to see people getting laid off anyway we need someone to bench it but if it had IPC gains they would clearly mention it even if it was single digit like crestmont only time will tell

They are keeping it quiet for Granite Rapids but come back to the original comment by Pat again "10-plus percentage change in the core..."

Otherwise, the move from Intel 4 to Intel 3 would have been a year delay with a minor process bump, meaning nothing real gained. It could have been Intel 4 GNR instead of EMR so it would have been an overall loss.

However, if the statement is true, then the year delay would have made sense.

The reddit comments by Intel folks from everywhere is saying it's gloomy. If you don't know who will be fired, why would you do good work? Truly this is the worst period for Intel, coming at a time when the macroeconomy has all indications for a crash too.

 

[mikk]

According to One Raichu they mixed LNC front end with RWC+

https://twitter.com/x/status/1828258538583732299

So what is Geddagod saying now?

Page 116 - Discussion - Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads

Page 116 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[igor_kavinski]

What's a good real-world test of latency impact? Maybe better than synthetics.

Maybe Total War: Warhammer III from here: https://www.techspot.com/review/2777-ddr5-vs-ddr4-gaming/



The game engine seems to be very sensitive to latency and DDR4-4000 CL16 gives first word latency of 8ns while the DDR5 there has first word latency of 9.44ns.

 

[511]

So what is Geddagod saying now?

Page 116 - Discussion - Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads

Page 116 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

That is 2023 post he was saying 🤣

 

[Philste]

So is Intel sandbagging LNCs IPC after all? They are claiming >20% ST performance over MTL on one of their Hot Chips Slides. Since LNLs clocks stayed basically the same, LNC IPC must be much better than 14% or am I missing something?

 

[DokiDoki]

What's a good real-world test of latency impact? Maybe better than synthetics.

CFD