Last year, Intel unveiled its strategic roadmap for artificial intelligence (AI) and accelerators, introducing a next-generation data center GPU codenamed "Falcon Shores". Positioned as the successor to the Gaudi 3, Falcon Shores boasts a multi-chip modular structure combined with a "scalable I/O design." This innovative GPU is set for release in 2025. To facilitate its production, Intel has partnered with TSMC, opting for a combination of 3nm and 5nm processes, enhanced by CoWoS advanced packaging technology, with mass production anticipated by the close of 2025.

At SC 2024, Intel expanded its strategic outlook by showcasing a new roadmap that introduces a chip codenamed "Jaguar Shores", the planned successor of Falcon Shores. Developed by Habana Labs, Jaguar Shores is envisioned as a versatile GPU tailored for AI training, inference, and high-performance computing (HPC) applications. The chip will be crafted using Intel's 18A process, which incorporates pioneering back-side power supply technology aimed at significantly enhancing energy efficiency.

Currently, Intel's lineup of AI chips is navigating several challenges, leading to some confusion in its product offerings. This complexity is reportedly influenced by internal strategic shifts and changes in customer demand. Notably, the projected shipments for the existing Gaudi 3 have been drastically slashed by over 30%. Gaudi 3 faces stiff competition from NVIDIA's Hopper-architecture-based H100, and further challenges loom with the impending release of products from the new Blackwell architecture. As the development cycle advances, Intel will phase out the Gaudi family, integrating its IP into Falcon Shores.

However, the Falcon Shores initiative has undergone significant changes over recent years, originally conceptualized as an XPU (CPU+GPU) and evolving into a pure GPU strategy. With the cancellation of the Ponte Vecchio successor, Intel’s pathway forward for Jaguar Shores remains open to various design possibilities, although the specific direction is yet to be disclosed. Observers speculate that maintaining a general-purpose GPU design will be pivotal moving forward.