重點資訊
Intel has announced its new GPU D-Series, formerly codenamed Crescent Island, which is specifically engineered for AI inference workloads. This hardware represents a strategic shift for the company as it attempts to address the needs of enterprise environments.
The D-Series utilizes Intel's Xe3P architecture to integrate up to 480 GB of LPDDR5X memory. It delivers 1.5 TB/s of memory bandwidth and 1.3 petaflops of FP8 compute performance, all contained within a 350-watt power envelope.
According to Intel VP of AI Products and GTM Anil Nanduri, the series aims to make high-capacity inference practical and economical to scale. The hardware is intended for diverse applications, including robotics, multi-tenant AI services, and other latency-sensitive tasks.
This release marks a departure from the previous Gaudi 3 chip, which featured 128 GB of HBM. By switching to LPDDR5X, Intel is prioritizing efficiency for inference over the training-focused approach of its earlier, less successful Gaudi and cancelled Falcon Shores projects.
