Intel Announces Inference-Optimized Xe3P Graphics Card with 160GB VRAM (www.phoronix.com)

Intel has formally unveiled "Crescent Island," an inference-focused data center GPU built on its next‑gen Xe3P (Celestial) micro‑architecture. The card is specified with a large 160 GB of LPDDR5x memory, multi‑data‑type support and is tuned for performance‑per‑watt, air‑cooled operation and lower cost — positioning it explicitly for LLM inference and "tokens‑as‑a‑service" workloads. Customer sampling is slated for H2 2026 with broader shipping likely later, and Intel provided few other hardware details today. The announcement matters because it signals Intel’s renewed push into the inference accelerator market against AMD’s Instinct MI450 and NVIDIA’s Vera Rubin, emphasizing energy efficiency and large model capacity rather than raw training throughput. Technically, 160 GB of LPDDR5x targets running larger models or batching more tokens in memory, while support for multiple data types aids inference flexibility. Intel also plans to use current Arc Pro B‑Series hardware and ongoing open‑source driver work (Project Battlematrix, Xe Linux drivers, Compute Runtime) to mature software support — a crucial factor for adoption. Overall, Crescent Island is intriguing but concrete performance, connectivity, and software ecosystem details remain sparse, and meaningful availability is still a year or more away.