Intel targets inference costs with memory-heavy D-series GPU

Intel unveiled its GPU D-series at the Open Compute Project Global Summit, targeting memory-intensive AI inference rather than competing primarily on peak compute. Previously codenamed Crescent Island and built on the Xe3P architecture, partner-built cards support up to 480GB of LPDDR5X memory. Intel’s own branded version carries 160GB. The design delivers 1.5TB/s of memory bandwidth and 1.3 petaflops of FP8 compute within a 350-watt envelope, using a standard, air-cooled PCIe card.

The strategy aims to reduce the need to split models and conversation caches across multiple accelerators, potentially lowering communication overhead and deployment costs. LPDDR5X offers cost and power advantages over HBM, but sacrifices bandwidth compared with high-end HBM-based accelerators. Intel says a single card can support models of roughly 300 billion parameters. That capacity claim has not been validated through independent performance testing, and does not establish practical serving throughput or latency.

Customer sampling is underway with server manufacturers, cloud operators and software vendors, but Intel has announced neither pricing nor a firm general-availability date. Its competitive prospects depend on measured serving costs, software compatibility and production availability. The narrower inference focus follows earlier Gaudi and Ponte Vecchio efforts that failed to significantly challenge Nvidia’s position.