Lunar Lake: Intel's Copilot+ Moment

After two weeks in Taipei, I'm finally back in San Diego having had the opportunity to contemplate much that has transpired and learned during Intel Technology Tour 2024 and Computex 2024 both held in Taipei, Taiwan. One thing is indisputable, the PC industry is at a pivotal moment with AI PC and a contentious Copilot+ PC standard declared at this year's Microsoft Build.

A gamble on AI PC design that could pay off at scale

Microsoft made it clear at Build 2024, Qualcomm's Snapdragon X Series of processors would christen Microsoft's Copilot+ PC initiative and the flagship devices from seven leading PC OEMs including Microsoft. This despite AMD having shipped their first AI PC processor over a year ago, and Intel launching their Core Ultra AI PC line with Meteor Lake six months ago.

It seems that Intel, much like AMD, got on the wrong side of the Copilot+ PC bet falling short of the 40+ NPU (Neural Processing Unit) TOPS (trillion operations per second) needed for PCs to qualify for Microsoft's on-device generative AI experience.

At CES 2024, I had the chance to preview Lunar Lake to get an early feel for what the new Intel AI PC processor would deliver a mere two weeks after Core Ultra's official launch with Meteor Lake. The Intel team did not disclose the details of the chip design but claimed it would give Snapdragon X Elite a run for its money including in the much-vaunted NPU department.

It's important to remember that at the time there was confusion (and continues to be) about what an AI PC is and how the standard was being defined. During CES 2024, MJ (Michelle Holthaus Johnson) alluded to Meteor Lake powering the Intel's first class of OEM devices that met Microsoft's standard for an AI PC satisfying the requirements of having a NPU, "comes with Copilot", and a Copilot branded key on the keyboard.

Then came the official 40+ TOPS of NPU compute requirement with the introduction of Copilot+ PC at Microsoft Build to confuse the AI PC matter even more.

At Intel Vision 2024, Intel suggested with little detail that Lunar Lake would pack over 100 "platform" (CPU + GPU + NPU) TOPS of AI compute which would outclass the 75 platform TOPS of Qualcomm's top-of-the-line Snapdragon X Elite.

It turns out that Intel was being modest. In front of an exclusive audience of media and industry analysts, MJ revealed that Lunar Lake would power in at up to 120 platform TOPS of AI compute with 48 of those TOPS coming from massively increased NPU compute over Meteor Lake's 11 TOPS.

It's notable that Lunar Lake will come to market with up to 67 TOPS of GPU AI compute for heavier AI workloads and the graphics performance that creators and gamers will appreciate in their AI PC. Lunar Lake will sport Intel's latest Xe2 GPU architecture with 8 Xe cores.

On paper, Lunar Lake is poised to get Intel on the Copilot+ PC wave to meet the Snapdragon X Series threat and power the company toward the 100 million Intel-based AI PCs that it and its OEM partners aim to sell by the end of 2025.

Designed to push x86 power efficiency into new frontiers

According to Intel, Lunar Lake is a complete rethinking of x86 PC processor making its architecture significantly different from Meteor Lake. The singular design principle for Lunar Lake was leading performance per watt to outclass front-running and emerging Arm-based platforms from Apple and Qualcomm.

The most interesting change in Lunar Lake is the new microarchitecture which departs from Meteor Lake's 6 P-core, 8 E-core, and 2 ultra-low-power E-core configuration with a dedicated SoC tile and GPU tile. Lunar Lake features a simplified 8-core hybrid (4 "Skymont" E-core + 4 "Lion Cove" P-core) design that introduces Memory on Package of up to 32 GB.

Most notably, Lunar Lake's Lion Cove P-core removes hyperthreading which reduces the area and power that this feature typically imposes on a chip design. The depreciation of hyperthreading could have the added and possibly unintended benefit of reducing Lunar Lake's vulnerability to speculative execution attacks. This could be a boon as Lunar Lake, much like Meteor Lake, is considered for applications beyond the PC such as for industrial IoT and edge infrastructure applications increasingly requiring trustable execution environments.

Another huge redesign is the compute tile that now incorporates the low-power island and deprecates the SoC tile as well as the standalone GPU tile characteristic of the Meteor Lake architecture. The GPU and NPU now reside on the compute tile in Lunar Lake manufactured on TSMC's N3B process technology. You may recall that Meteor Lake's compute tile was based on Intel 4 while the SoC tile and GPU tile leveraged TSMC's N6 and N5 processes respectively. All this results in a significant simplification and process technology uplift.

An AI processor for AI computing beyond the NPU

Intel made some interesting bets on the Lunar Lake design. I call it an AI compute heavy design with heft on the NPU and GPU. In terms of CPU cores configuration, Lunar Lake looks a lot more like Apple's M3 and is likely outgunned in multicore performance by the Snapdragon X Elite's twelve all-performance core design based on Qualcomm's Oryon CPU cores. We will know more about how these processors stack up to each other once devices get in the hands of benchmarking specialists who are itching to run their own tests and compares.

Lunar Lake's design seems to reflect Intel's view on how the AI PC will play out and evolve. Clearly, Intel sees the present and future of AI PC computing as an XPU thing rather than a NPU thing as Microsoft's Copilot+ PC marketing might suggest.

According to Intel's own research, ISVs will be considering a diversity of AI engines to optimally serve the diversity of their AI applications and architectures. As a relatively new core type in PC computing, the NPU faces the headwind of developer awareness and learning curves, but Intel sees increased use of NPU as more ISVs learn to capitalize on its differentiation as an AI engine.

The overall case for a multi-engine approach is apparent for AI applications that involve RAG (Retrieval Augmented Generation) architectures involving a collection of vector databases, retrievers, and chatbot middleware. These applications can involve a range of AI and non-AI operations that may not optimally execute on NPU.

Intel's AI PC volume play in the era of Microsoft Copilot+ PC

There is no denying it. Intel is in the uncomfortable situation of having ceded the Copilot+ PC pole position to Qualcomm. This is a first for Intel which has always been tied to the hip with Microsoft. It's also a first for Qualcomm which has found itself in the enviable position of being Microsoft's prime and sole partner in the launch of Copilot+ PC with the first devices hitting shelves on June 18.

We expect that Lunar Lake will kick in around September, just in time for back-to-school and the holiday shopping seasons around the globe. By then, Qualcomm and AMD will have had Copilot+ PC devices in the market for two months or more. AMD's new Ryzen AI 300 processor touting 50 NPU TOPS will be arriving in Copilot+ PC devices sometime in July.

At the end of the day, the AI PC will be fought on the mobile front. Battery life is what will matter. With NPU TOPS largely and arguably equal, the biggest question is whether Intel will deliver on a performant AI PC processor in Lunar Lake that disproves the performance per watt legend of Arm-based processors while getting a leg up on AMD in the process.

If Intel can do this, they will be able to leverage their curiously disregarded scale and go-to-market muscle to minimize hits on their coveted PC market share and drive the volumes in AI PCs that Microsoft needs to scale their Copilot ambition.

Subscribe to the neXt Curve Insights monthly newsletter to be notified when the next newsletter is published. Go to www.next-curve.com to be added to our mailing list. You will also be notified when we publish new research notes and media content.

Follow us on LinkedIn, Twitter, and YouTube.

This material may not be copied, reproduced, or modified in whole or in part for any purpose except with express written permission or license from an authorized representative of neXt Curve. In addition to such written permission or license to copy, reproduce, or modify this document in whole or part, an acknowledgement of the authors of the document and all applicable portions of the copyright notice must be clearly referenced.

© 2024 neXt Curve. All Rights Reserved