Intel Announces Inference-Optimized Xe3P Graphics Card with 160GB VRAM

Summary

Back during the Intel Tech Tour in Arizona, Intel teased a new inference-optimized enterprise GPU would be announced soon. This new product would feature enhanced memory, bandwidth, and enterprise-level AI inference capabilities. Today the embargo expires on talking about this new GPU offering. When Intel was teasing this new inference-optimized GPU a few weeks back in Arizona it sounded like Intel may have had an unexpected trick up its sleeves. What's being announced today is indeed a new enterprise GPU for AI that is interesting from a technology perspective, but it's not shipping until at least H2'2026. So while there was hope that perhaps Intel had managed to innovate some interesting Battlemage / BMG-G31 part for AI or the like with lots of vRAM, what's being announced is a next-gen part but one that is at least one year away still. This new graphics card is codenamed Crescent Island and is built on their next-gen Xe3P Celestial micro-architecture. Xe3P will be optimized around performance-per-Watt and Crescent Island will feature 160GB of LPDDR5x memory to allow for plenty of space for large language models (LLMs). Intel's embargoed announcement also notes that Crescent Island will feature support for a variety of different data types and be an "ideal" solution for tokens-as-a-service providers and inference use cases. In addition to being optimized around performance-per-Watt, Crescent Island will also be air cooled and cost-optimized. Intel is currently working on refining their open-source software stack for Crescent Island via using current-generation Arc Pro B-Series GPUs. Intel's announcement notes that customer sampling of this new data center GPU will begin in the second half of 2026. No official release timeframe was provided if they also hope to squeeze it out next year or if (more than likely) it will actually ship more broadly in 2027 but just noting their customer sampling for H2'2026 in the embargoed news release. No slides or prototype images o...