Intel "Lunar Lake" a Ground-Up Architecture Designed to Dominate the 15W Segment

[btarunr]

Intel faces its stiffest long-term challenge not from AMD in the client desktop and server segments, but from the likes of Apple and Qualcomm in the ultraportable notebook SoC market. The company is looking to meet the challenge head-on with not just its current Hybrid architecture-based "Rocket Lake-U" processors, but the upcoming "Meteor Lake" and "Arrow Lake" architectures, which disaggregate the processor into chiplets build on various foundry nodes, including ones that are external to Intel. The move here is to keep Intel ahead of the curve with Moore's Law buckling. The real ace up Intel's sleeves is "Lunar Lake."

Slated for "beyond 2024" (which could mean 2025 or later), "Lunar Lake" is reportedly a fresh ground-up design for not just the SoC, but also the CPU microarchitecture, with a design focus on performance/Watt, and a focus on mobile devices. This, according to a statement to Dr Ian Cutress. Intel has already brought Hybrid CPU cores to the PC, and with Arm partners loading their SoCs with three or more kinds of CPU cores; Intel could possibly give "Lunar Lake" even more [kinds of] Hybrid processing capabilities. Another hint on the direction in which Intel is heading comes from unlikely quarters—"Sapphire Rapids." Intel's latest enterprise processors attempt to overcome the CPU core-count deficit to AMD, by incorporating various on-die accelerators—these are fixed-function hardware that accelerate specific kinds of enterprise workloads. Intel could give us another lick about "Lunar Lake" in its January 26 Q4-2022 Financial Disclosures Day.



View at TechPowerUp Main Site | Source

 

[natr0n]

I read disintegrated

 

[Hyderz]

do you think they could make a laptop that can last a full day - average use

 

[watzupken]

Until I see the actual product, all these are just finger in the air talks. Building sometime ground up to focus on low power devices is a good start. But looking at Intel's track record over last 5 years or so, they have always been pushing performance at the expense of power requirement. So I will be interested to see what they will deliver given the intense competition.

In any case, I feel that x86 processors are not really suited for the ultra low power segment. ARM based SOC seems like a better fit for such lower power requirements and for battery life. So I am looking forward to see what Qualcomm can bring to the table since my experience with Apple's M1 Macbook Air was really good.

 

[randomUser]

At this point it doesn't seem like anyone could beat M1 even by 2025.

 

[john_]

15W base, 250W turbo?

 

[bug]

When companies shift focus to products that are years away, that's a sign their current offering won't get anything exciting by then.
But it happens, from time to time you have to let go of one thing and start working on the next.

 

[Dirt Chip]

So much can go wrong and delay it to 2026+.
This mix tile\lithography is great but each different tile add complexity and so possible fail point of many sorts.

 

[usiname]

At this point it doesn't seem like anyone could beat M1 even by 2025.

No M1, M1 pro was destroyed long ago. No one need this useless OS limited toys, for somehting more than media device. Ouch, I forgot to mention overpriced, RAM and SSD limited

 

[konga]

No M1, M1 pro was destroyed long ago. No one need this useless OS limited toys, for somehting more than media device. Ouch, I forgot to mention overpriced, RAM and SSD limited
View attachment 279528

The 6900HS is 10% faster on average if you remove the AES results while consuming over twice as much power. I would not call this "destroyed."

Say what you want about the OS and software experience (I'm not an apple guy for sure), but the silicon is still undefeated when it comes to power efficiency. Which is what this is about.

 

[hs4]

current Hybrid architecture-based "Rocket Lake-U" processors

????

I believe the purpose of the "Royal Core" as it was once called was to clean up the waste from Skylake's continued buildup and to create a leaner design that delivers modern IPC.

As for piling three different cores, it's possible, if only for the possibility of adding more cores for standby time; Intel officially stated at Hot Chips 30 (2018) that the Cove-mont hybrid is Big-bigger, not Big-LITTLE. P-E combination is the same as the Prime-big combination in ARM for Android, and the Intel hybrid lacks a standby time LITTLE core.

 

[ARF]

There won't be any "20A" in this year, nor "18A" later because Intel is stuck with its 10nm and its ++++++++ versions

Class-action lawsuits should follow for misleading investor and public information.

 

[konga]

There won't be any "20A" in this year, nor "18A" later because Intel is stuck with its 10nm and its ++++++++ versions

Class-action lawsuits should follow for misleading investor and public information.

Misleading investors can result in a lawsuit, but proving that a statement was misleading can be a pretty high bar. Them failing to meet an overly ambitious goal is not enough. As the plaintiff, you would have to prove, with actual evidence, that Intel never had any intention of meeting the goals they laid out from the start. Good luck with that.

 

[hs4]

There won't be any "20A" in this year, nor "18A" later because Intel is stuck with its 10nm and its ++++++++ versions

Class-action lawsuits should follow for misleading investor and public information.

Intel is getting more than even performance with 5nm at 10nm++++. Could it be that the design is that much better?

At least video editors using MacOS are disappointed with the M1 Ultra's lack of performance in Da Vinci Resolve and Photoshop (and even Final Cut Pro developed by the Apple themselves.) Threadripper was also popular among them, but unfortunately AMD has abandoned it for Socket sTRX4; the 13900K was reasonably popular because it could deliver the same level of performance as the M1 Ultra even with 60W PL2, and the same level of performance per watt as the 3970X.

 

[TheoneandonlyMrK]

do you think they could make a laptop that can last a full day - average use

They need to because the competition is bringing just that.
To be fair though even now they(Intel) do ok, my I5 work laptop gets used at least 7 hrs for work without a power lead.

 

[Karti]

There won't be any "20A" in this year, nor "18A" later because Intel is stuck with its 10nm and its ++++++++ versions

Class-action lawsuits should follow for misleading investor and public information.

remember how intel last 14nm+++++++++ (hahaha look i am funny) was almost as same as TSMC 7nm under the laser microscope..?

or how Intel already had 10nm CPUs but people spammed "still only 14nm+++++++" because consumer desktop series was still at 14nm? Will give you a hint, look at 11th gen mobile segment, Intel already had 10nm then on some models

these comments of "place random number, add nm and spam alot of + symbols" are just boring

 

[TheoneandonlyMrK]

remember how intel last 14nm+++++++++ (hahaha look i am funny) was almost as same as TSMC 7nm under the laser microscope..?

or how Intel already had 10nm CPUs but people spammed "still only 14nm+++++++" because consumer desktop series was still at 14nm? Will give you a hint, look at 11th gen mobile segment, Intel already had 10nm then on some models

these comments of "place random number, add nm and spam alot of + symbols" are just boring

As are PR releases of for example sapphire rabbids for 3/10 years before release.
Continuously slipping release schedules.
And promises of Angstrom grade parts years before release.

But we're here again with the lofty PR anoucment without any real facts and absolutely no parts to show.

 

[Vya Domus]

but the silicon is still undefeated when it comes to power efficiency.

Except that's not true because it's dumb to use "undefeated power efficiency" as a general blanket statement.

You know what's actually the most power efficient silicon ? Probably something like this : https://www.amd.com/en/products/cpu/amd-epyc-9654, you can be sure there are plenty of test cases where 96 cores at 360W are going to annihilate Apple silicon at power efficiency.

 

[usiname]

The 6900HS is 10% faster on average if you remove the AES results while consuming over twice as much power. I would not call this "destroyed."

Say what you want about the OS and software experience (I'm not an apple guy for sure), but the silicon is still undefeated when it comes to power efficiency. Which is what this is about.

And 20% faster when you remove geektrash from the chart. With tunning the ryzen will still be faster with half the power, something that cannot be achieved from the locked toys. And if we speak for efficiency, 35w for 12k points 7950x can make double of this with 35w. Next time when you open your mouth, chech the price, just $2000 for base configuration wih 8core cpu and $2500 for the 10 core version, double of the price of 6900hs laptop with discrete gpu. Imagine to spend so much for such trash

 

[ARF]

these comments of "place random number, add nm and spam alot of + symbols" are just boring

What random number? Intel renamed its 10nm to something like "Intel 7" that hasn't even got "nm" metric besides it.

Also, "20A" is TSMC N3, while "18A" is in fact TSMC N2.
Years will pass before you see that in your Core i9...

 

[JustBenching]

At this point it doesn't seem like anyone could beat M1 even by 2025.

Im pretty sure both intel and amd smack the m1 in both efficiency and performance.

 

[zlobby]

do you think they could make a laptop that can last a full day - average use

Full day means 24h?

 

[Gica]

Misleading investors can result in a lawsuit, but proving that a statement was misleading can be a pretty high bar. Them failing to meet an overly ambitious goal is not enough. As the plaintiff, you would have to prove, with actual evidence, that Intel never had any intention of meeting the goals they laid out from the start. Good luck with that.

The current template "xx nm" is completely outdated and work is being done on another, more realistic standard. It's no wonder that Intel changed the name, too many measure with the "xx nm" meter, when, in reality, in an Intel 14nm mm2 you had the same density of transistors as in a TSMC 10nm.

 

[Minus Infinity]

Lunar Lake might be very impressive but it won't see the light of day until 2026 IMO. Arrow Lake is late 2024 at best since Meteor Lake got pushed back until end of the year and is now only mobile.

I think Arrow Lake will be a nice improvement on efficiency, but yeah Lunar Lake is the big one.

By that time Zen 6 will be out too and Intel will face stiff competition. Zen 4 offers far more performance per Watt below 125W and in laptop Dragon range HX will slaughter Raptor Lake HX in the 45-65W segment.

 

[hs4]

Below 100nm, scaling-dependent changes in physical properties (imagine large bodies of water flowing, but small drop sticking to the walls against gravity) make it impossible to shrink in a isotropic manner, and have been addressed by changing the shape of elements and cells. Planar MOSFETs have become tri-gate/FinFETs, soon to become gate-all-around/RibbonFETs. Intel4 has only shrunk vertically in cells relative to intel7.

In such a situation, the "nm" metric, which assumes a isotropic shrinkage, has no physical meaning. It has been customary to double the transistor density in one generation, and to reduce the "X" in "X nm" to 1/√2. However, Intel rounded 2.4x to the nearest 2x, while TSMC rounded 1.7x to the nearest 2x. That gave roughly equal steps between Intel's two generations and TSMC's three generations.

Intel is aware that small steps are more reliable than one large stride in the latest node, and they said at last year's conference that it will be a series of small steps in the future.

Lunar Lake might be very impressive but it won't see the light of day until 2026 IMO. Arrow Lake is late 2024 at best since Meteor Lake got pushed back until end of the year and is now only mobile.

I think Arrow Lake will be a nice improvement on efficiency, but yeah Lunar Lake is the big one.

By that time Zen 6 will be out too and Intel will face stiff competition. Zen 4 offers far more performance per Watt below 125W and in laptop Dragon range HX will slaughter Raptor Lake HX in the 45-65W segment.

Intel has only a small quantity of EUV machines in 2023 because of the delay in ordering EUV machines from ASML, but there are quite a few for delivery in 2024.

As a result, mass production for Intel4 will take place at only one location, Fab34, while the 20A Fab is being built at four locations in Arizona and Ohio, and according to the general civil schedule, will be completed and receive EUV machines in 2024.