At Hot Chips 2026, Intel provided new details on Crescent Island, its upcoming graphics card aimed at data centers and workstations. The card is built on the new Xe3P GPU architecture and is designed as a 350W, air-cooled, PCIe-format accelerator compatible with a range of system configurations. Unlike traditional GPUs, Crescent Island contains no graphics-output hardware; the silicon is built purely for AI inference workloads, including long-context and agentic tasks.
The chip is organized into four Compute Slices, each containing eight Xe3P cores, for a total of 32 cores, 256 Vector Engines, and 256 XMX (matrix) engines, backed by 32 MB of unified L2 cache. Each core includes 512 KB of L1/SLM cache and large register files, described as 1 MB per core. A key architectural change from the standard Xe3 design is the XMX engine pipeline, which moves from a four-deep systolic array to a 16-deep design in Xe3P. The Vector Engines add FP8 and FP4 support alongside extended math and FP64 capability, and the architecture overall supports data types ranging from FP4 to FP64.
On memory, Intel-branded Crescent Island cards will ship with 160 GB of LPDDR5X, while Intel is allowing ODM partners to build versions with up to 480 GB of LPDDR5X for maximum capacity. Intel says using LPDDR5X instead of HBM lowers both power draw and cost while still enabling a densely packed memory channel design for strong bandwidth. For comparison, competing accelerators expected around the same time include AMD's Instinct MI450X, offering up to 432 GB of HBM4, and Nvidia's Vera Rubin, offering up to 288 GB of HBM4.
Intel frames Crescent Island for PC, workstation, and data-center edge deployments where fast AI inference is needed. As an example, running a trillion-parameter model such as Kimi K3 locally would require roughly 2 TB of aggregate system memory; four Crescent Island GPUs at the maximum 480 GB configuration could reach that total in a single workstation, though such a memory configuration would be costly given current DRAM prices.
On the software side, Intel says Crescent Island will have day-zero support for major AI frameworks, including vLLM, SGLang, llm-d, and Nvidia Dynamo, and will support a broad range of model types such as language, reasoning, multi-modal, and diffusion models. Intel also highlights KV-cache-aware routing and offloading for scalable inference, along with the ability to run agent workloads across heterogeneous infrastructure without changes to agent code. The Xe3P architecture underlying Crescent Island is also expected to appear in Intel's next client GPU lineup, the Arc C-Series, reflecting its scalability from client integrated graphics to data-center accelerators. Intel is targeting customer sampling of Crescent Island in the second half of 2026.