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	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.)

 

[DrMrLordX]

The game developer would incorporate a small LLM that is trained specifically on the areas the NPCs are supposed to be able to speak about. This model would be shipped with the game and would use whatever supported accelerators the machine has (GPU, NPU, etc.).

The developer isn't giving up control since they control the model and they can train and fine tune it as needed.

For these kind of purposes, small and specialized models make much more sense than the massive do everything LLMs we often see today. Such small models can perform quite well on modest hardware.

If they could nail the thing down properly, it would make consumer NPUs (Meteor Lake, Arrow Lake for example) more useful (or at least make it easier to sell products featuring NPUs to gamers). Methinks game devs would be better off farming out a lot of that work to other developers who specialize in AI.

@Thunder 57

Hallucinations et al would be a problem. I would think that most existing game devs wouldn't be prepared for that sort of thing. The last thing you would want is for the LLM to make statements related to in-game lore that happened to be subtly inaccurate in some way that wouldn't fit the NPC.

 

[Hulk]

Point that should be seen is that in Intel hybrid cpu designs HT gain in best case scenario isn't 30% but at best 10%. It should be plain obvious that SMT should be dropped from P-cores and instead focus targeting best possible 1T-speed with those P-cores - and let highly-threaded workloads scale with E-cores. Intel should be able to compensate that 10% MT perf drop from SMT easily resulting better design in all performance points.

I agree that focusing the P cores on ST and using the E's for MT is the future. I've stated it many times.

But as is often the case when discussing things in forums that was not my original point. My original point was that currently Raptor Cove "enjoys" a 31% improvement in IPC when running Cinebench R23 MT with HT enabled.

I'm exhausted with the mis-reading of my posts but I'll go through it one last time. If you are going to reply to this post all I ask is that you please read this entire post.

The max stock speeds of a 14900K all-core are 5.7/4.4. CB R23 MT P score is ~23,300 and for the E's 18,600 for a total score of 41,900. Shut off HT and the total score drops to 36,400.

That is a 15% performance loss for the entire chip w/o HT.

Now, I realize that running a 14900K 5.7/4/4 is quite a task so I backed that 41,900 score down to 40,000 as a comparison point for ARL 8+16. Many reviews have achieved 40,000 points without custom loops.

Finally, my point is this. When you take into account the clock speed regression of ARL along with the lack of HT and the latency penalty for going with tiles, making up that 15% is going to be a tough nut to crack and I will be impressed if ARL 8+16 is able to score 40,000 points in CB R23 MT.

That's it, that's is all I'm saying. I hope Intel can do it but I see it as a big mountain to climb. I agree with the direction Intel is going removing HT from the P's. I also agree that there will be a significant increase in ST IPC for the P's and E's and that in most software ARL will be more performant than the 14900K. But in "ridiculously" multi-threaded apps the 14900K will still be a high bar for ARL to reach.

The great thing about this is all will be revealed soon enough. It won't be a point of contention or a mystery forever.

 

[ondma]

Point that should be seen is that in Intel hybrid cpu designs HT gain in best case scenario isn't 30% but at best 10%. It should be plain obvious that SMT should be dropped from P-cores and instead focus targeting best possible 1T-speed with those P-cores - and let highly-threaded workloads scale with E-cores. Intel should be able to compensate that 10% MT perf drop from SMT easily resulting better design in all performance points.

I haven't tried it myself, but I think the gain from HT is more than 10%, even in the hybrid design i9 tiers with lots of E cores (more like 15-20%???). And, you are giving up a more significant benefit in the lower performance tiers where there are less E cores.

 

[ondma]

I haven't tried it myself, but I think the gain from HT is more than 10%, even in the hybrid design i9 tiers with lots of E cores (more like 15-20%???). And, you are giving up a more significant benefit in the lower performance tiers where there are less E cores.

Edit: Hulk has the data already. I didnt see it. From his data it looks like the gain from HT is 15% or slightly over.

 

[igor_kavinski]

From his data it looks like the gain from HT is 15% or slightly over.

If you compare Geekbench File Compression MT HT vs. No HT score, it can be anywhere from 20% to 30% depending on architecture.


Old article of Ivy Bridge HT performance impact on lossless compression workloads: https://www.tomshardware.com/reviews/winrar-winzip-7-zip-magicrar,3436-11.html





IF Raptor Lake HT has 30% compression performance advantage, that goes poof with ARL.

Some workloads will miss HT.

 

[naukkis]

I agree that focusing the P cores on ST and using the E's for MT is the future. I've stated it many times.

But as is often the case when discussing things in forums that was not my original point. My original point was that currently Raptor Cove "enjoys" a 31% improvement in IPC when running Cinebench R23 MT with HT enabled.

I'm exhausted with the mis-reading of my posts but I'll go through it one last time. If you are going to reply to this post all I ask is that you please read this entire post.

The max stock speeds of a 14900K all-core are 5.7/4.4. CB R23 MT P score is ~23,300 and for the E's 18,600 for a total score of 41,900. Shut off HT and the total score drops to 36,400.

That is a 15% performance loss for the entire chip w/o HT.

Problem is that Intel did extract everything they could from chip - and gone too far. They become unstable. Big part of that is that SMT - removing SMT in chip design will decrease chips hot spots by simplying critical path routings. Have to actually wonder if anybody is actually tested what those Intel chips actual stable performance figures are - if SMT is one of problems which push chip stability out of edge and simply disabling SMT could bring stability back at high frequencies.

 

[Hulk]

Problem is that Intel did extract everything they could from chip - and gone too far. They become unstable. Big part of that is that SMT - removing SMT in chip design will decrease chips hot spots by simplying critical path routings. Have to actually wonder if anybody is actually tested what those Intel chips actual stable performance figures are - if SMT is one of problems which push chip stability out of edge and simply disabling SMT could bring stability back at high frequencies.

Exactly right. I have a 14900K and run it w/o HT, 5.5/4.3, air cooled, always under 1.3V, 75C max temp and it's 100% stable. I could put it more but for insignificant gains in the daily work loads.

Here's the thing. I like the K parts as they are. Intel gives you specs on what they can do with serious cooling and tweaking. If you don't want to deal with that then back off the frequency a bit or buy the non K parts.

The problem I have with the K parts is the wide variability in the v/f curve among CPU's. Most KS parts for example will do 6.2GHz with over 1.5V but some golden samples will do it just under 1.4V. That's a big variation for parts that cost the same. But hey that's part of the KS "lottery" I guess.

 

[igor_kavinski]

Have to actually wonder if anybody is actually tested what those Intel chips actual stable performance figures are - if SMT is one of problems which push chip stability out of edge and simply disabling SMT could bring stability back at high frequencies.

I hope someone sends their defective 14th gen chip to igorslab for investigation.

 

[Hulk]

I hope someone sends their defective 14th gen chip to igorslab for investigation.

I can run less voltage without HT for the same stability.

 

[igor_kavinski]

I can run less voltage without HT for the same stability.

But it would be interesting to see if a crashing and possibly degraded CPU's stability can be restored and if so, at what settings and what is the price (performance downgrade) to salvage a CPU like that.

 

[Abwx]

I can run less voltage without HT for the same stability.

Of course since HT will require 20-30% higher peak current for the cores.

Hence there s 18-23% less voltage drops in the supply voltage rails that get through the cores layout, so where 1.5V was required to have the final voltage at the end of the conductors it s now no more than 1.25-1.3V that is required for an equivalent stability, that s basic Ohm s law.

 

[AMDK11]

Skylake - SunnyCove
micro-ops(decode + uop cache) from 11 to 11 +0%
Dispatch/Rename from 4 to 5 +25%
execution ports from 8 to 10 +25%
With 2xFP/ALU + 2xALU, 1xS/D + 3xAGU
for 2xFP/ALU + 2xALU, 2xS/D + 4xAGU
IPC average +18%

SunnyCove - GoldenCove
micro-ops(decode + uop cache) from 11 to 14 +27%
Dispatch/Rname from 5 to 6 +20%
execution ports from 10 to 12 +20%
With 2xFP/ALU + 2xALU, 2xS/D + 4xAGU
for 3xFP/ALU + 2xALU, 2xS/D + 5xAGU
FPU+ALU from 4 to 5 +25%
IPC average +19%

GoldenCove - LionCove
micro-ops(decode + uop cache) from 14 to 24 +71.4%
Dispatch/Rename from 6 to 8 +33.3%
execution ports from 12 to 18 +50%
With 3xFP/ALU + 2xALU, 2xS/D + 5xAGU
up to 4xFPU, 6xALU, 2xS/D + 6xAGU
FPU+ALU from 5 to 10 +100%
IPC increase +??%

 

[H433x0n]

Skylake - SunnyCove
micro-ops(decode + uop cache) from 11 to 11 +0%
Dispatch/Rename from 4 to 5 +25%
execution ports from 8 to 10 +25%
With 2xFP/ALU + 2xALU, 1xS/D + 3xAGU
for 2xFP/ALU + 2xALU, 2xS/D + 4xAGU
IPC average +18%

SunnyCove - GoldenCove
micro-ops(decode + uop cache) from 11 to 14 +27%
Dispatch/Rname from 5 to 6 +20%
execution ports from 10 to 12 +20%
With 2xFP/ALU + 2xALU, 2xS/D + 4xAGU
for 3xFP/ALU + 2xALU, 2xS/D + 5xAGU
FPU+ALU from 4 to 5 +25%
IPC average +19%

GoldenCove - LionCove
micro-ops(decode + uop cache) from 14 to 24 +71.4%
Dispatch/Rename from 6 to 8 +33.3%
execution ports from 12 to 18 +50%
With 3xFP/ALU + 2xALU, 2xS/D + 5xAGU
up to 4xFPU, 6xALU, 2xS/D + 6xAGU
FPU+ALU from 5 to 10 +100%
IPC increase +??%

You're missing the details of it adding new levels of cache and the moderate increase in cache overall. In an ideal scenario Lion Cove could maybe be +20%.

However, It’s attached to a cursed IOSF tiled config so it’s practically guaranteed to be <20%. It has its memory controller on a separate die, it’s going to most likely have a decently large L3$ penalty and a memory latency penalty of +30ns. Redwood Cove core for all intents and purposes should perform on par with Raptor Cove clock for clock but it’s overhead has it losing to RPL-H by sometimes up to 10% at isoclocks in synthetic benchmarks.

ARL-S only real shot is that its SoC tile is significantly better than the version in MTL-H. We already know it removes the LP ecores, the IOSF frequency can be overclocked. There won’t be power constraints so the ring will probably run at 4.5ghz instead of 3.2-3.4ghz like MTL. It’s Intel though and good things aren’t allowed to happen for them so I wouldn’t expect much.

 

[Joe NYC]

as well as UE5 somehow incorporate AI acceleration (increased fps at lower power consumption as NPC AI calculations are offloaded onto the NPU).

I strongly doubt this is going to happen in time frame relevant to ARL

 

[igor_kavinski]

I strongly doubt this is going to happen it time frame relevant to ARL

They already have MTL NPU to test out the feature if it's been under development. I mean, think about it. Why would Intel specifically market ARL as "the first gaming CPU with AI" ? What the hell do gamers care about productivity on their gaming CPU? Gotta be something games related.

A thought: What if it's a global XeSS filter offloaded to the NPU?

 

[maddie]

They already have MTL NPU to test out the feature if it's been under development. I mean, think about it. Why would Intel specifically market ARL as "the first gaming CPU with AI" ? What the hell do gamers care about productivity on their gaming CPU? Gotta be something games related.

A thought: What if it's a global XeSS filter offloaded to the NPU?

Come on now. Don't you realize that Intel is throwing out buzzwords everywhere about everything.

 

[igor_kavinski]

Come on now. Don't you realize that Intel is throwing out buzzwords everywhere about everything.

I don't know. Something doesn't seem right. Intel must have something special in mind to excite gamers about AI otherwise why even bother? They could've chosen a different tagline like "First gaming CPU with an NPU" but saying AI means it could've applications in gaming.

I forgot that AMD beat them to the punch with the 8700G: https://www.pcgamer.com/worlds-firs...-desktop-pcs-in-the-second-half-of-this-year/

So now they have to say something like "First unlocked gaming CPU with AI accelerator" or something.

 

[reggie_fils_aime]

They already have MTL NPU to test out the feature if it's been under development. I mean, think about it. Why would Intel specifically market ARL as "the first gaming CPU with AI" ? What the hell do gamers care about productivity on their gaming CPU? Gotta be something games related.

A thought: What if it's a global XeSS filter offloaded to the NPU?

An ancillary thought: What if Intel also has no idea why 99% would have a single reason to want an NPU, and is merely selling based on present-day hype?

 

[dullard]

An ancillary thought: What if Intel also has no idea why 99% would have a single reason to want an NPU, and is merely selling based on present-day hype?

So two possible thoughts come to mind.

Either:
1) Over 100,000 Intel employees and not one can find even a single list out of thousands of lists of actual useful prompts that many people will want: https://copilot.cloud.microsoft/en-US/prompts

or

2) You have nothing useful to say in an Intel thread.

I'm really torn as to which of those two options is more likely to be true.

 

[reggie_fils_aime]

So two possible thoughts come to mind.

Either:
1) Over 100,000 Intel employees and not one can find even a single list out of thousands of lists of actual useful prompts that many people will want: https://copilot.cloud.microsoft/en-US/prompts

or

2) You have nothing useful to say in an Intel thread.

I'm really torn as to which of those two options is more likely to be true.

I'm just very cool on the whole "AI" thing, because 99.9% of the stuff that's marketed towards normal people are really basic tasks that anyone literate could complete in a few minutes. The present implementation of LLMs are somewhere between tacky and counterproductive, and mainly seem to exist for kids to cheat on their homework and office clerks to pump out peppier form letters. It's impressive, but not in a way that makes me say "I need this now".

 

[dullard]

I'm just very cool on the whole "AI" thing, because 99.9% of the stuff that's marketed towards normal people are really basic tasks that anyone literate could complete in a few minutes. The present implementation of LLMs are somewhere between tacky and counterproductive, and mainly seem to exist for kids to cheat on their homework and office clerks to pump out peppier form letters. It's impressive, but not in a way that makes me say "I need this now".

LLMs aren't going to run on an NPU--at least not any time soon. You are looking at the wrong types of AI. Look at small language models instead--ones with specific applications for use.

 

[reggie_fils_aime]

Sorry

LLMs aren't going to run on an NPU--at least not any time soon. You are looking at the wrong types of AI. Look at small language models instead--ones with specific applications for use.

Sorry, I was using LLM imprecisely as an analogue for "you can talk at it".

 

[dullard]

Sorry

Sorry, I was using LLM imprecisely as an analogue for "you can talk at it".

I too don't think that talking with a computer is very useful. It is interesting, and might make games a bit more realistic and less repetitive. But, talking is still very limited use. What about non-talking tasks?

• Find a trend between these two data sets.
• Find all the photos of me hiking with my son in the 1960s.
• What did my wife say about XYZ?
• What were my assigned tasks from the meeting that I missed?
• Remove this stranger from my favorite photo.
• Is there a defect in the items on the conveyor belt?
• Which plumping connection do I need to accomplish ABC based off of this photo?
• What will this look like in my house if I buy it?
• Etc.

These are not tasks that "anyone literate could complete in a few minutes". Some people maybe, but most not.

Is AI great for everyone right now? No, it is in its infancy. But anyone who just outright thinks that 99% would not have a single use for AI just lacks any form of imagination. AI on an NPU is then for the subset of use cases where you don't want to put everything online (uploading all photos to do the photo search, latency issues during manufacturing, security issues, business secrecy issues with data trends, etc).

 

[reggie_fils_aime]

All of that seems fine and good, and would certainly negate the need for trained human supervision. I'm perhaps a little jaded on that whole aspect because this is all work I've been trained to do throughout my adult life: finding trends in data, editing photos, concise and clear communications... That's a broadcasting and journalism degree, in so many words. I guess I'm not the target audience, lol.

 

[Joe NYC]

They already have MTL NPU to test out the feature if it's been under development. I mean, think about it. Why would Intel specifically market ARL as "the first gaming CPU with AI" ? What the hell do gamers care about productivity on their gaming CPU? Gotta be something games related.

A thought: What if it's a global XeSS filter offloaded to the NPU?

Gaming PCs desktops are not particularly power constrained, since they are typically plugged to the wall. They already have the capabilities of the NPU in the GPU.

As far as behavior of NPC characters, NPCs will most likely be governed by the CPU for some time to come. Far more CPU resources are being used to display an NPC character than to decide what the NPC character does.

When it comes to better hardware utilization, for better "AI", just using more CPU cores (that have already been there for years) more efficiently would be one area.

Then there is AVX, AVX512. There are more gaming PCs / CPUs that have AVX512 vs NPU.