Intel AI Everywhere: Architectural Leaps Set the New Foundation
Intel’s AI Everywhere event showcased the formal launch of Intel Core Ultra and their next-generation server CPU. The event highlighted a turning point for Intel regarding product architecture and manufacturing.
- Intel emphasized its scale as one of the key reasons it will play a significant role in bringing AI to the edge in PC devices
- Next-generation architecture in Core Ultra sets a new era for Intel SoC design
- Server roadmap continues to mature with chiplet architectures and dedicated AI solutions like Gaudi
- 4 nodes in 5 years execution remains on track, and Intel is likely to deliver on being the #2 foundry by 2030
A point was made by Michelle Johnston Holthaus, Intel’s EVP & GM of Client Computing Group that I think needs to be emphasized. During the official launch of Intel Core Ultra she said “Intel Core Ultra is our largest shift in architecture in 40 years.” Thinking about Intel’s client roadmap going forward, this point has to be emphasized. Intel Core Ultra represents a completely new architecture and SoC design approach for Intel where they leverage a chiplet architecture with advanced packaging for the first time in a client SoC. Another key point for Core Ultra is it is the first to use Foveros 3D packaging. It is also the first product built on any node that wasn’t already in production when the 5 nodes in 4 years initiative began. This new architectural approach will be a key milestone moment for Intel as this first generation Core Ultra sets the stage for Intel client product architecture innovation for the foreseeable future.
When it comes to AI, we believe we will continue to see a doubling down of AI accelerators as a part of next-generation architecture design. This is again evident with Intel Core Ultra as it is the first client SoC from Intel with an NPU, which will be a continued technical conversation around dedicated blocks on client silicon to accelerate AI tasks. We are working on a dedicated report on the NPU which will be out early next year.
Lastly, on the client story, it is hard to ignore Intel’s scale. Over the next two years, between 1st and 2nd generation Core Ultra, Intel will ship “conservatively” ~100m PCs into the market which will be capable of advanced AI running locally on-device. Meaning that from a Windows/x86 developer standpoint, Intel silicon will have the largest addressable installed base for software developers and enterprises looking to do more local AI software.
From a data-center standpoint, Intel continues to mature its advanced chipset architecture and advanced packaging techniques with Emerald Rapids. Performance gains look to be significant gen on gen with Emerald Rapids, and the key positioning around AI into this new architecture is Intel has built AI accelerators in every one of the processor’s 64 cores, which gives users 42% higher inference performance versus the previous generation creating a more cost-effective way to run generative AI models. It is again important to highlight the significant TAM expansion ahead for the data center. Intel’s continued execution and roadmap here will remain competitive in both the typical data center refresh and the billions of dollars being additive to infrastructure silicon spending each year. To highlight the growth potential here Morgan Stanley estimates total AI server market (including hardware, memory/storage and computing) growing from ~$55B in 2023 to ~$100B by 2025 (35% CY23-25 CAGR), adding ~ $45B of annual enterprise hardware spend over the next 2 years.
Lastly, at each event CEO Pat Gelsinger continues to state that the 5 nodes in 4 years initiative remains on track, and his confidence in Intel regaining process leadership with 18A can’t be understated. Gelsinger is not the kind of person who would continue to state such confidence in this execution if he had any doubt about delivering. I think this should be one of the continued biggest stories for the Intel turnaround thesis, as well as the high likelihood that Intel will execute and become the #2 foundry by revenue by 2030. There is an AI silicon manufacturing story brewing here as well, as Intel’s advanced packaging technologies are in many ways more competitive than TSMCs and this point may very well become more evident as semiconductor designers begin to adapt their design strategies with AI silicon in mind and need a more robust manufacturing and packaging solution. From numerous semiconductor industry insider conversations I’ve had there is a belief that Intel will be well positioned not just for advanced process manufacturing but as a foundry with specific expertise in AI silicon manufacturing and packaging.
Summary of Event News
- The team is progressing successfully with its PC and production strategies, leading to the recent introduction of the Intel Core Ultra Processor (Meteor Lake). They are on schedule with their planned releases both in the near term and looking ahead. Intel’s Core Ultra processor has been unveiled, showcasing the Intel 4 process node with EUV lithography and Foveros packaging technology. This processor adopts a revolutionary multiple-chip die architecture which combines both the CPU and a neural processing unit within a unified system. This advancement is a significant leap in strengthening AI software development capabilities in the PC sphere. Intel anticipates integrating 300 AI-accelerated features into PCs and aims to collaborate with over 100 software allies, including prominent names like Adobe, Dolby, and Zoom.
- Intel’s server product trajectory is well-aligned with the introduction of their next-gen server CPU (Emerald Rapids) on Intel 7. Their operational execution on product and technology roadmaps is evidenced by the rollout of the Emerald Rapids-based 5th Gen Xeon platform (Intel 7). Boasting a notable 21% average performance improvement over its predecessor, the 5th Gen Xeon processor has incorporated AI accelerators into each of its 64 cores. This enhancement delivers a 42% boost in inference performance, offering a more efficient solution for running advanced generative AI models like LAMA 2. The servers powered by the 5th Gen Xeon are set to become widely accessible from the first quarter of 2024 through leading providers such as Cisco, Dell, and HPE. Additionally, the team has publicly showcased its Gaudi 3 accelerator for the first time, with availability slated for 2024.