Intel returns to AI with Crescent Island, this time focusing on inference

Under Pat Gelsinger, Intel tried to take on Nvidia and AMD with its Gaudi GPUs, but sales were weak and the company cancelled the planned Falcon Shores successor. Rather than abandoning AI, Intel is now pursuing a different angle: hardware aimed at inference workloads.

The Financial Times spoke with Kevork Kechichian, Intel’s general manager of its data centre group, and at Computex Crescent Island was shown with additional details beyond last year’s announcement. Unlike Gaudi, Crescent Island is built for running already‑trained models, so it doesn’t require large amounts of pricey High Bandwidth Memory, can be air‑cooled instead of using complex liquid systems, and relies on cheaper LPDDR5X — a design that could help ease the global memory shortage if the market shifts toward inference.

Nvidia is also targeting inference and has teamed up with Groq to develop a chip that blends a language accelerator with its Rubin platform.

intel, crescent island, inference, gaudi, falcon shores, pat gelsinger, kevork kechichian, lpddr5x, hbm, nvidia