Inside Intel’s Lunar Lake: A Promise That Became a Problem

[Fred Chen]

郭明錤 (Ming-Chi Kuo) Nov 5, 2024

Intel has recently announced that after Lunar Lake (LNL), it will discontinue integrating DRAM into CPU packaging. While this news has caught public attention recently, people working in the PC industry have known for at least six months — according to Intel’s roadmap, upcoming products like Arrow Lake, Nova Lake, Raptor Lake refresh, Twin Lake, Panther Lake, and Wildcat Lake will not follow LNL’s packaging approach.

Two main motivations drove the LNL project:
1. Competing with Apple Silicon: Following the rise in MacBook market share thanks to Apple Silicon, Intel aimed to prove that the x86 architecture could achieve similar performance and battery life.
2. Responding to Microsoft’s Move: Intel learned Microsoft’s new Surface lineup (2Q24) would use Qualcomm chips with 45 TOPS of AI computing power, so Intel planned to launch a competing product to counter this.

To achieve the above goals, LNL made three key decisions:
1. Integrating DRAM into the package.
2. Specifying certain components, such as designating Renesas as the exclusive supplier of power management chips.
3. Boosting NPU computing power to 48 TOPS, slightly above Qualcomm’s X Elite/Plus at 45 TOPS.

Notably, among Intel’s 2024–2025 processors, only LNL exceeds Microsoft’s defined AI PC requirement of 40 TOPS. This seemingly unusual spec planning was because LNL was positioned to compete with Qualcomm. However, Intel likely didn’t know Microsoft would set the AI PC requirements of 40 TOPS.

Interestingly, this unusual LNL specification became a brief marketing advantage for Intel this year, giving them at least one positive story — their AI PC solution — amid a stream of negative news. But this was pure luck. If Intel had learned about Microsoft’s 40 TOPS requirement earlier, they wouldn’t have planned Arrow Lake with only 36 TOPS total — less than the AI PC requirement and LNL’s 48 TOPS.

Intel claims LNL failed because integrated memory hurt their gross margins, but the real story is different:
1. PC manufacturers weren’t interested because they lost flexibility in choosing components, which hurt profits.
2. The cost structure is unfavorable because Intel’s bargaining power for DRAM supply is much lower than Apple’s, and it must rely on TSMC’s production.
3. Customers don’t want to pay more for LNL because AI PC applications are immature.

LNL’s failure indicates that Intel’s challenges extend beyond foundry technology limitations. The company’s deeper issues lie in product planning, as also evidenced by AMD’s continued gains in the conventional server market. While manufacturing technology gaps often dominate discussions, Intel’s fundamental challenge might be organizational, leading to flawed product decisions.

Inside Intel’s Lunar Lake: A Promise That Became a Problem

Intel has recently announced that after Lunar Lake (LNL), it will discontinue integrating DRAM into CPU packaging. While this news has…

medium.com

 

[Xebec]

I think Arrow Lake mobile will overlap Lunar Lake's segment to offer a DRAMless option with similar performance characteristics to LNL. (ARL also being N3, same architectures, and "strong" iGPU IIRC). Though I'm unsure what TOPS ARL mobile will be capable of.

It makes sense Apple has the most DRAM bartaining power of any OEM since they also bundle with iPhone and iPad, not just Mac.

 

[Maximus]

Intel claims LNL failed because integrated memory hurt their gross margins, but the real story is different:
1. PC manufacturers weren’t interested because they lost flexibility in choosing components, which hurt profits.
2. The cost structure is unfavorable because Intel’s bargaining power for DRAM supply is much lower than Apple’s, and it must rely on TSMC’s production.
3. Customers don’t want to pay more for LNL because AI PC applications are immature.

When I first saw Pat's 4Y5N plan, I felt it was too risky to assume Intel would regain its strong market position solely by reclaiming process leadership. We've already seen products like the Arc GPU, Ponte Vecchio, Arrow Lake, and Lunar Lake, where the compute tiles are manufactured by TSMC. Their issues are not related to the manufacturing process.

 

[Xebec]

When I first saw Pat's 4Y5N plan, I felt it was too risky to assume Intel would regain its strong market position solely by reclaiming process leadership. We've already seen products like the Arc GPU, Ponte Vecchio, Arrow Lake, and Lunar Lake, where the compute tiles are manufactured by TSMC. Their issues are not related to the manufacturing process.

I'm not sure they're mutually exclusive.

I completely agree Intel definitely has major design and execution issues. Arrow and Lunar Lake are the clearest examples (assuming N3B is actually significantly more performant at higher frequencies than N4 as used by AMD).

But Intel having access to strong(er) internal processes and packaging could still provide balance sheet benefits (if they have the volume) and performance benefits that can help overcome (mildly) weaker products.

 

[siliconbruh999]

When I first saw Pat's 4Y5N plan, I felt it was too risky to assume Intel would regain its strong market position solely by reclaiming process leadership. We've already seen products like the Arc GPU, Ponte Vecchio, Arrow Lake, and Lunar Lake, where the compute tiles are manufactured by TSMC. Their issues are not related to the manufacturing process.

Ponte vecchio is a cursed product alongside SPR ARL used same flawed logic as MTL so it got punished LNL was a good product but the cost structure made it bad it has been a mix of bad decision and bad design choices

 

[blueone]

2. The cost structure is unfavorable because Intel’s bargaining power for DRAM supply is much lower than Apple’s, and it must rely on TSMC’s production.

I don't understand this statement. TSMC doesn't make DRAM, or am I mistaken?

 

[siliconbruh999]

I don't understand this statement. TSMC doesn't make DRAM, or am I mistaken?

Well the memory is on same package as CPU intel assembles the whole package at it's fab/assembly site so they have to deal with the DRAM purchase and OEM don't like it and neither Intel due to associated cost for Intel higher cost for oem one less thing to cheap out on

 

[blueone]

Well the memory is on same package as CPU intel assembles the whole package at it's fab/assembly site so they have to deal with the DRAM purchase and OEM don't like it and neither Intel due to associated cost for Intel higher cost for oem one less thing to cheap out on

I know all of this, and predicted OEMs wouldn't like various things about in-package DRAM, but I still don't understand the statement that Intel must rely on TSMC's production. I can't make sense out of it.

 

[nghanayem]

I don't understand this statement. TSMC doesn't make DRAM, or am I mistaken?

I read it as two separate ideas due to the ", and". That and it also doesn't make sense any other way since TSMC doesn't make DRAM or package LNL.

 

[blueone]

I read it as two separate ideas due to the ", and". That and it also doesn't make sense any other way since TSMC doesn't make DRAM or package LNL.

You might be right, but that sentence would still be D- english grammar.

 

[Fred Chen]

I don't understand this statement. TSMC doesn't make DRAM, or am I mistaken?

TSMC doesn't supply the DRAM that is bundled into the packaging. It just meant relying on TSMC is unfavorable for Intel's cost structure.

 

[hist78]

TSMC doesn't supply the DRAM that is bundled into the packaging. It just meant relying on TSMC is unfavorable for Intel's cost structure.

At Intel's scale, I believe it has the ability to get one of the best prices in sourcing LPDDR5X memory for Lunar Lake. But did Intel's Foveros advanced packaging used by the Lunar Lake cost too much that offsets all other cost savings? Can Foveros' hight cost justify the benefits of performance gains and superior features of Lunar Lake?

 

[Xebec]

At Intel's scale, I believe it has the ability to get one of the best prices in sourcing LPDDR5X memory for Lunar Lake. But did Intel's Foveros advanced packaging used by the Lunar Lake cost too much that offsets all other cost savings? Can Foveros' hight cost justify the benefits of performance gains and superior features of Lunar Lake?

I'm also curious why we don't seem to see any significant latency improvements with on package DRAM on Intel.

 

[blueone]

I'm also curious why we don't seem to see any significant latency improvements with on package DRAM on Intel.

I wasn't anticipating latency improvements, but in-package DRAM does enable the use of higher clock frequencies on the LPDDR5 channels at lower power consumption levels than if the DRAM chips were on DIMMs or even soldered to a circuit board. So you get much higher throughput. In Lunar Lake's case, Intel used the highest possible transfer rate of 8.533GT/sec. Even the super-duper ASUS TUF A16 Ryzen 9 gaming laptop is only running DDR5-4800 DIMMs. The same is true for the Dell Intel i9 laptops. Once you're on a circuit board you it blows your loss budget, so I'm told by my analog friends.

 

[VCT]

What if Lunar Lake laptop become a homerun in the market? I like it a lot and tell people around me go buy it while it's available.

It's a good product, consumers like it but Intel think profit margin is more important than a good product.

 

[Brady]

Intel increased orders for Lunar Lake so I’d say that for the state Intel is in this would be called a “good problem”.

Intel To Reportedly Outsource More “Arrow Lake” Orders To TSMC; Team Blue No Longer Confident In Its Foundry Business

edit: a good problem for their Products division, not so much for Foundry

 

[blueone]

What if Lunar Lake laptop become a homerun in the market? I like it a lot and tell people around me go buy it while it's available.

It's a good product, consumers like it but Intel think profit margin is more important than a good product.

I think the "problem" is more likely their major laptop OEMs. A friend who works at a big PC client OEM said they hate Lunar Lake. Greater percent of system value goes to Intel, lack of configuration flexibility, products from different OEMs will only be differentiated by the display quality, keyboard and touchpad quality, and packaging, and the latter will look like Apple's. And he said they're not interested in competing with Apple on those points, because Apple has superior margins and pricing power. And MacBooks are only about 8% of the market by unit volume.

I predicted these reactions before the conversation, but I think the laptop OEMs won't have a choice. I think a lot of customers will like Lunar Lake laptops better. I also think the game is only at the top of the first inning. Now OEMs, dump your bloatware on Windows that make your products seem so lame and annoying.

 

[OneEng]

郭明錤 (Ming-Chi Kuo) Nov 5, 2024

Intel has recently announced that after Lunar Lake (LNL), it will discontinue integrating DRAM into CPU packaging. While this news has caught public attention recently, people working in the PC industry have known for at least six months — according to Intel’s roadmap, upcoming products like Arrow Lake, Nova Lake, Raptor Lake refresh, Twin Lake, Panther Lake, and Wildcat Lake will not follow LNL’s packaging approach.

Two main motivations drove the LNL project:
1. Competing with Apple Silicon: Following the rise in MacBook market share thanks to Apple Silicon, Intel aimed to prove that the x86 architecture could achieve similar performance and battery life.
2. Responding to Microsoft’s Move: Intel learned Microsoft’s new Surface lineup (2Q24) would use Qualcomm chips with 45 TOPS of AI computing power, so Intel planned to launch a competing product to counter this.

To achieve the above goals, LNL made three key decisions:
1. Integrating DRAM into the package.
2. Specifying certain components, such as designating Renesas as the exclusive supplier of power management chips.
3. Boosting NPU computing power to 48 TOPS, slightly above Qualcomm’s X Elite/Plus at 45 TOPS.

Notably, among Intel’s 2024–2025 processors, only LNL exceeds Microsoft’s defined AI PC requirement of 40 TOPS. This seemingly unusual spec planning was because LNL was positioned to compete with Qualcomm. However, Intel likely didn’t know Microsoft would set the AI PC requirements of 40 TOPS.

Interestingly, this unusual LNL specification became a brief marketing advantage for Intel this year, giving them at least one positive story — their AI PC solution — amid a stream of negative news. But this was pure luck. If Intel had learned about Microsoft’s 40 TOPS requirement earlier, they wouldn’t have planned Arrow Lake with only 36 TOPS total — less than the AI PC requirement and LNL’s 48 TOPS.

Intel claims LNL failed because integrated memory hurt their gross margins, but the real story is different:
1. PC manufacturers weren’t interested because they lost flexibility in choosing components, which hurt profits.
2. The cost structure is unfavorable because Intel’s bargaining power for DRAM supply is much lower than Apple’s, and it must rely on TSMC’s production.
3. Customers don’t want to pay more for LNL because AI PC applications are immature.

LNL’s failure indicates that Intel’s challenges extend beyond foundry technology limitations. The company’s deeper issues lie in product planning, as also evidenced by AMD’s continued gains in the conventional server market. While manufacturing technology gaps often dominate discussions, Intel’s fundamental challenge might be organizational, leading to flawed product decisions.

Inside Intel’s Lunar Lake: A Promise That Became a Problem

Intel has recently announced that after Lunar Lake (LNL), it will discontinue integrating DRAM into CPU packaging. While this news has…

medium.com

Intel has a great number of issues.

DESIGN

Lion Cove (P cores) barely eeks out a single thread IPC advantage over Zen 5 while taking up a larger die space on a more dense process.

If that weren't bad enough, Lion Cove also suffers from a lack of AVX 512 and SMT. In heavily threaded applications Zen 5 just tears it a new one. Now, Skymont (Intel E core) is a much better story .... but only in Desktop and laptop (more on this in a moment), but lacks high clock speeds needed to compete with Zen 5c in higher performance situations. Still, it is small and efficient (unlike Lion Cove). For desktop and laptop MT applications, lots of Skymont cores does the trick.

Multiple designs of P and E cores and dis-similar architectures are expensive to produce and validate and make a mess for the OS scheduler.

Business Case

Intel has been a machine in the CPU business largely because for a few decades, they managed to maintain a 1 to 2 node process advantage over all other competitors. They managed such a dominant position that they were ALSO able to add a steady stream of new instructions to x86 that AMD could not get into production for another design cycle after Intel released it. Intel also used its superior products to squeeze AMD in any market that AMD was doing OK in by pricing its direct competing products below AMD's profit line while making up the money in DC and high end applications.

Fast forward to today.....

New nodes are exponentially more expensive while market prices are level (at best). Intel has fallen BEHIND TSMC in process technology. AMD has taken the initiative in chiplets and advanced multi-chip designs. They have also created a single core technology that can be used across all market segments in x86 (Desktop, Laptop, DC).

Intel is unable to compete in the DC market (although Granite Rapids did get them within 40% of AMD Turin based on the latest full reviews) where the market is growing rapidly and price pressures on processors is low.

But wait you say! Intel now has Arrow Lake to base their next gen Clearwater Forest on and its quite good Skymont E cores.

This seems like a tough sell to me. Skymont does not have SMT. While Skymont is very space efficient and power efficient, it is bested by Zen 5c by better than 40% per core in highly threaded DC applications. So .... just add more Skymont cores you say? Sure, but ...... in DC, it is frequently the case where the software licensing FAR exceeds the cost of the server hardware. Many software packages are based on the number of cores for their licensing. This means that using lots of NON-SMT single cores vs fewer SMT single cores is going to be very very expensive to the end user.

Now, perhaps Intel will surprise me and Skymont will have a variant with SMT .... but I doubt it. Processors need to be designed from the ground up for such a design and it looks like Skymont was designed for very efficient (space and power) single core performance.

Intel's 18A certainly looks attractive though. Still, with densities being around that of N3B (someone correct me if I am off here), I am not sure if the "juice is worth the squeeze". Intel has spent a huge amount of money chasing foundry supremacy in order to regain the model of their past success. Even IF they manage to achieve this with 18A, the process will only be used for CWF in 2025, and IMHO, CWF appeal will be limited due to DC software licensing issues it will face. Intel will be dumping more money into 16A which will be attempting to bring up the ASML’s NXE 5000 series EUV scanners (High NA) to my understanding. Again, grossly expensive for use on a grossly expensive highly multi-pass process node (GAA with BSPD).

With such a business model, how is it possible for Intel to see profits?

Still, an earlier poster had it right. Intel's product management has completely failed the company. They have not produced the right products at the right price. Furthermore, Intel seems hell bent on doing what they have been doing for the last 30 years even though it is clear that this path leads to financial ruin.

In a more philosophical question, I wonder how the world of CPU's progresses in the future when it can't rely on double the transistor budget every 18 months?

 

[Paul2]

At Intel's scale, I believe it has the ability to get one of the best prices in sourcing LPDDR5X memory for Lunar Lake. But did Intel's Foveros advanced packaging used by the Lunar Lake cost too much that offsets all other cost savings? Can Foveros' hight cost justify the benefits of performance gains and superior features of Lunar Lake?

The whole point is that Intel gets monster LPDDR deal, and pockets the tiny, little premium that would've been going to the memory supplier, or a memory distributor.

Many smaller OEM would welcome this technical solution: no need to design high speed lines to memory for laptop PCB, nor tinker with the power solution by yourself.

You can use much cheaper PCB, with fewer layers, not risk that the few customisations that you have to do will backfire after manufacturing.