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.





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.)

 

[Tigerick]

Very true. It's hilarious to see one set of people saying MTL is as bad as bulldozer (or even worse) and the other set of people saying it's the awesomest ever!

Last 24 hours has been better than watching sports!


Yep. MTL's RWC was a safe bet for Intel. ARL is gonna be their first true moonshot in client (with LNC & 20A). I think ARL's CPU tile will get more than 50% PPW gain over MTL's CPU tile. Their first true Apple competitior.

Let me stop your unrealistic dream. Ian has recently interviewed Ann Keheller about 18A, here is what she claimed about 18A compared to Intel 4:-

Intel 3 : 18% performance per watt gain (from Intel 4)
Intel 20A : 10% ppw gain (from Intel 3)
Intel 18A : 10% ppw gain (from Intel 20A)

Let's assume performance/watt figure is 1, then compared to Intel 4:-

Intel 3: 1+18% = 1.18 ~18% better PPW
intel 20A: 1.18+10% = 1.298 ~ 30% better PPW
Intel 18A: 1.298+10% = 1.4278 ~ 43% better PPW

There you have:- Compared to Intel 4, 20A at best provides 30% better PPW and 18A should provide 43% better PPW. Intel 18A without High-NA seems offer limited PPW improvement, maybe that's the reasons only Panther Lake with 4P+8E in the roadmap.

And no, 18A is not comparable to TSMC's N2. In fact, N2 might be used by Intel for upcoming Nova Lake cause Intel still has not updated the process node that's comparable to N2....

 

[Hitman928]

Did you write this a few days ago?

Or this little conversation a little further back:

Take with a grain of salt. Not sure if this is ST or MT...
View attachment 84943

That's 1t power.

Rubbish. 22W @ 4.8GHz 1T would make it the biggest joke of the year. Not even Intel is capable of something silly like that.

For starters, MTL is expected to be extremely power efficient. Intel even said that the Intel 4 process node has been heavily optimized for power efficiency. And RWC is not exactly the same as Raptor Cove. RWC is heavily power optimized for MTL. The entire tile architecture in MTL is also designed with extreme power efficiency in mind. I'm think MTL's power efficiency is gonna be out of the world.

I wuldn't be surprised if MTL mobile can comfortably match the performance of RPL mobile at just half the power. Maybe even less!

Hence Ultra. Their supposed-to-be Centrino moment!

Definitely met expectations, just the wrong person's expectations. Funny how goalposts can change so quickly.

 

[Meteor Late]

Let me stop your unrealistic dream. Ian has recently interviewed Ann Keheller about 18A, here is what she claimed about 18A compared to Intel 4:-

Intel 3 : 18% performance per watt gain (from Intel 4)
Intel 20A : 10% ppw gain (from Intel 3)
Intel 18A : 10% ppw gain (from Intel 20A)

Let's assume performance/watt figure is 1, then compared to Intel 4:-

Intel 3: 1+18% = 1.18 ~18% better PPW
intel 20A: 1.18+10% = 1.298 ~ 30% better PPW
Intel 18A: 1.298+10% = 1.4278 ~ 43% better PPW

There you have:- Compared to Intel 4, 20A at best provides 30% better PPW and 18A should provide 43% better PPW. Intel 18A without High-NA seems offer limited PPW improvement, maybe that's the reasons only Panther Lake with 4P+8E in the roadmap.

And no, 18A is not comparable to TSMC's N2. In fact, N2 might be used by Intel for upcoming Nova Lake cause Intel still has not updated the process node that's comparable to N2....

Besides, Intel 4 was supposed to bring 20% PPW uplift over Intel 7, reality is <10%, and it's not like we are talking about the higher part of the curve, Intel showed a curve about Intel 4 where 20% efficiency uplift was achieved around 2-3GHz, Meteor Lake is operating at those frequencies in MT workloads. Why would we believe Intel's numbers about future nodes?

 

[igor_kavinski]

Why would we believe Intel's numbers about future nodes?

You raise a good point. Love your username

 

[SmokSmog]

So what is the real ST clock of "5ghz"😉 Core Ultra 7 under heavy load? Cuz 4,6ghz ADL beats it in R23 ST.
Intel 4 is a dumpster fire

 

[tamz_msc]

So what is the real ST clock of "5ghz"😉 Core Ultra 7 under heavy load? Cuz 4,6ghz ADL beats it in R23 ST.
Intel 4 is a dumpster fire

Define heavy load first. y-cruncher 1T?

 

[igor_kavinski]

Intel 4 is a dumpster fire

Maybe late 2024 steppings of MTL will improve with Intel 4+ refined node which they might call Intel Enhanced 4 node. Then release MTL Refresh to cover the SKUs that won't be served by Looney Lake.

 

[igor_kavinski]

Define heavy load first. y-cruncher 1T?

Excel using countifs(). Yes, with an s.

 

[SiliconFly]

Intel 3 : 18% performance per watt gain (from Intel 4)
Intel 20A : 10% ppw gain (from Intel 3)
Intel 18A : 10% ppw gain (from Intel 20A)

According to Intel's published data, it's actually:

Intel 3 : 18% performance per watt gain (from Intel 4)
Intel 20A : 15% ppw gain (from Intel 3)
Intel 18A : 10% ppw gain (from Intel 20A)

So, from Intel 4 to Intel 20A, its a total of 15% ppw gain over 18%. 100% -> (18%) -> 118% -> (15%) -> 136%. Thats around 36% gain total.

And this 36% is just the node gains alone, not including any power efficiency gains from LNC's new architecture. In effect, 50% ppw gains for ARL's 20A CPU tile is something we can easily expect. Probably more if LNC has a more efficient design that what we suspect.

 

[SiliconFly]

Did you write this a few days ago?

Of course I did.

Well, we can always wait and watch how MTL sales fares in the coming months compared to competition.

 

[Meteor Late]

According to Intel's published data, it's actually:

Intel 3 : 18% performance per watt gain (from Intel 4)
Intel 20A : 15% ppw gain (from Intel 3)
Intel 18A : 10% ppw gain (from Intel 20A)

So, from Intel 4 to Intel 20A, its a total of 15% ppw gain over 18%. 100% -> (18%) -> 118% -> (15%) -> 136%. Thats around 36% gain total.

And this 36% is just the node gains alone, not including any power efficiency gains from LNC's new architecture. In effect, 50% ppw gains for ARL's 20A CPU tile is something we can easily expect. Probably more if LNC has a more efficient design that what we suspect.

Yeah, "easily".
Meteor Lake was also supposed to get at minimum 20% gain from "just the node" and we've seen how it has gone...

 

[eek2121]

Very true. It's hilarious to see one set of people saying MTL is as bad as bulldozer (or even worse) and the other set of people saying it's the awesomest ever!

Last 24 hours has been better than watching sports!


Yep. MTL's RWC was a safe bet for Intel. ARL is gonna be their first true moonshot in client (with LNC & 20A). I think ARL's CPU tile will get more than 50% PPW gain over MTL's CPU tile. Their first true Apple competitior.

Clearly they know nothing about Bulldozer.

Intel themselves don't think that.

Yes they do.

Me hyping? I think you are very confused. MTL has met all my expectations. I very clearly said MTL's performance will be similar to that of RPL (even a bit less in certain workloads) & it's efficiency will be a lot higher compared to RPL in very light loads due to LP E cores. Everything what I said has come true. Kindly read my words clearly before accusing me of hyping.

I don’t recall a single person hyping it. MTL landed exactly within my own expectations: more efficient Raptor Lake.

Assuming Intel gets the software issues surrounding LPe cores worked out, it will be quite a compelling platform.

Let me stop your unrealistic dream. Ian has recently interviewed Ann Keheller about 18A, here is what she claimed about 18A compared to Intel 4:-

Intel 3 : 18% performance per watt gain (from Intel 4)
Intel 20A : 10% ppw gain (from Intel 3)
Intel 18A : 10% ppw gain (from Intel 20A)

Let's assume performance/watt figure is 1, then compared to Intel 4:-

Intel 3: 1+18% = 1.18 ~18% better PPW
intel 20A: 1.18+10% = 1.298 ~ 30% better PPW
Intel 18A: 1.298+10% = 1.4278 ~ 43% better PPW

There you have:- Compared to Intel 4, 20A at best provides 30% better PPW and 18A should provide 43% better PPW. Intel 18A without High-NA seems offer limited PPW improvement, maybe that's the reasons only Panther Lake with 4P+8E in the roadmap.

And no, 18A is not comparable to TSMC's N2. In fact, N2 might be used by Intel for upcoming Nova Lake cause Intel still has not updated the process node that's comparable to N2....

18A is absolutely comparable to N2.

Besides, Intel 4 was supposed to bring 20% PPW uplift over Intel 7, reality is <10%, and it's not like we are talking about the higher part of the curve, Intel showed a curve about Intel 4 where 20% efficiency uplift was achieved around 2-3GHz, Meteor Lake is operating at those frequencies in MT workloads. Why would we believe Intel's numbers about future nodes?

You are comparing early benchmarks of an imperfect chip to a mature chip and blaming the process for the difference. Not very objective are we?

Username checks out I guess.

 

[Meteor Late]

How is it performing according to expectations? Meteor Lake has lower efficiency than Phoenix even with higher effective core count, when power limits are enforced properly, as evidenced by the few reviews that do this out there.
The consensus was that this is an "efficiency focused" architecture and not performance focused, yes, but the efficiency increase is small, still worse than the competition even with core count advantage (which boosts MT efficiency) and there is ST regression.

 

[SiliconFly]

Yeah, "easily".
Intel 7 to Intel 4 was also 20% gain from "just the node" and we've seen how it has gone...

It's not apples-to-apples comparison. Intel 7+ (ESF2)'s RPL is based on high power UHP cells that run stable and are slightly more efficient at (and only at) higher clocks than Intel 4's HD cells that don't clock well and has a higher power draw at similar higher clocks. Intel 4 is an outdated node by today's standards. 20A in 2024 isn't. Starting from Intel 4 to Intel 20A, it's all HD cells for client. so, the gains are more straight forward and in line with published data.

 

[Meteor Late]

It's not apples-to-apples comparison. Intel 7+ (ESF2)'s RPL is based on high power UHP cells that run stable and are slightly more efficient at (and only at) higher clocks than Intel 4's HD cells that don't clock well and has a higher power draw at similar higher clocks. Intel 4 is an outdated node by today's standards. 20A in 2024 isn't. Starting from Intel 4 to Intel 20A, it's all HD cells for client. so, the gains are more straight forward and in line with published data.

Alder Lake and Raptor Lake actually have similar efficiency in MT workloads at various power levels:

 

[SiliconFly]

Just to clarify, I like MTL a lot. But I'm not a big fan of the RWC CPU tile, as it's based on an outdated & power hungry architecture that shouldn't exist now. And based on a Intel 4 node thats already a bit outdated by today's standards.

The only amazing part of MTL is its 3D stacked tiled design, a better tGPU and the interesting new LP E cores which significantly improves efficiency in light workloads. RWC CPU tile is just a crutch and a safe bet Intel made to ensure timely delivery. Not a fan of the CPU tile.

So overall, MTL is definitely more efficient than previous gen. Better graphics. Not much performance improvement. Amazing new architecture. And stands up to competition. Thats it.

 

[SiliconFly]

Alder Lake and Raptor Lake actually have similar efficiency in MT workloads at various power levels:

...Starting from Intel 4 to Intel 20A, it's all HD cells for client. so, the gains are more straight forward...

 

[Meteor Late]

BTW, there is a huge middle area between heavily MT workloads and idle or very light tasks.
Most of the time, what matters most is single core and few-core power consumption and efficiency. That's why Apple is king in "real" efficiency, yeah it won't look that good if you compare standard M3 against 8-core 16-thread or 16-core 22-thread in MT workloads, but at what matters most, it is the best hands down, when equalized for number of cores, the difference is HUGE.

 

[SiliconFly]

Thats why Intel should ditch the outdated RWC soon and move to LNC. RWC still goes up to 5 GHz for peak performance which isn't good for power efficiency. Imagine hitting similar peak performance at a significantly lower clock (like say 3.5 to 4 GHz) which provides huge power efficiency gains. Apple has already done it. And I'm hoping LNC does something similar.

It's very much possible considering one unverified LNC leak which says LNC doesn't have much performance gains over previous previous gen. If true, then LNC's primary focus is just power efficiency alone, possibly hitting peak performance at a lower clock like Apple SoC. Otherwise Intel wouldn't bother to completely switch from RWC to LNC. Just a theory.

 

[adroc_thurston]

Definitely met expectations, just the wrong person's expectations. Funny how goalposts can change so quickly.

I hate being right but that's what I usually am.

 

[Hulk]

There's some weird cherry-picking going on. The Raptor Lake part is +9% vs the MTL part even though it's only 4% faster in clocks. Doesn't that suggest an IPC regression versus Raptor Lake?

View attachment 90454

Monolithic designs are generally more efficient than chiplets when architecture is the same. Even if the cores are the same between Raptor and Meteor there is a penalty for the tiles. So while technically IPC is the same, throughput, which is real world performance is greater for Raptor.

I have a suspicion that the Redwood cove architecture does have improvements that Intel is not telling us about because they are "hiding" the reduced throughput of going to the tiled design with the slightly improved Redwood Cove core.

I also think it is telling that the top of the line H part doesn't come until next year. It would suggest they are binning for that part.

 

[Saylick]

Monolithic designs are generally more efficient than chiplets when architecture is the same. Even if the cores are the same between Raptor and Meteor there is a penalty for the tiles. So while technically IPC is the same, throughput, which is real world performance is greater for Raptor.

I have a suspicion that the Redwood cove architecture does have improvements that Intel is not telling us about because they are "hiding" the reduced throughput of going to the tiled design with the slightly improved Redwood Cove core.

I also think it is telling that the top of the line H part doesn't come until next year. It would suggest they are binning for that part.

Looks like RWC is just a minor tweak of GLC. "New core architecture" my butt, and this is coming straight off Intel's accusations of snake oil...

https://twitter.com/x/status/1735641034812481788

 

[adroc_thurston]

Looks like RWC is just a minor tweak of GLC. "New core architecture" my butt, and this is coming straight off Intel's accusations of snake oil...

It's a tick.
Technically new march, but very minor changes overall.

 

[igor_kavinski]

march

uarch

Don't force people to think too hard!

 

[SiliconFly]

Looks like RWC is just a minor tweak of GLC. "New core architecture" my butt, and this is coming straight off Intel's accusations of snake oil...

RWC is just a rehash. Good news is, this is the end-of-the-line for all the old fat power-hungry P cores (GLC, RPC, RWC, etc). LNC should be better than RWC.

And more importantly, many days ago I mentioned that Intel marketing published the snake oil stuff to preempt AMD from taking potshots at MTL's crappy CPU tile. Well, no one believed me that time. It all adds up beautifully now!

Here's the excerpt:
"...AMD has the tendency to crash Intel's party by saying stuff like Intel's cores are monolithic... ...nodes are bad... ...Zen architecture is better than..."

"...as far as MTL is concerned... ...RWC which is just a rehash of RPC... ...is Intel's achilles heel. And probably the singular issue with MTL. And they know AMD will somehow bring it up to rain on the MTL parade. So, it appears they chose a preemptive strike. A carefully orchestrated attack on AMD to ensure they doesn't get too cozy taking potshots..."

Sadly, it has worked so far for Intel.