Intel has revealed technical details for its GPU D series accelerator, codenamed Crescent Island, alongside high density Xeon server racks designed for agentic AI workloads. Built on the Xe3P architecture, the air cooled PCIe card supports up to 480 GB of LPDDR5X memory to run massive language models without liquid cooling. Hardware sampling is underway with server partners, though retail pricing remains unannounced.
The D series accelerator runs on the Xe3P architecture, omitting display outputs to dedicate all silicon area to inference calculations. Physical hardware includes 32 Xe3P cores, 256 vector engines, and 256 redesigned XMX engines featuring 16 deep systolic processing. The card delivers 1.3 PetaFLOPS of FP8 compute power, supporting formats from FP4 to FP64. A media engine with 4 encoders and 4 decoders manages video processing. Memory capacity reaches 160 GB on Intel branded models, while hardware partners can configure up to 480 GB of LPDDR5X memory across 1.5 TB/s of bandwidth. Operating within a 350W power envelope, the card fits into standard PCIe slots and functions on conventional server air cooling.
A single D series accelerator runs models scaling up to 300,000,000,000 parameters. Linking 8 cards into a single cluster expands capacity to handle models exceeding 1,000,000,000,000 parameters. Anil Nanduri, vice president of AI products at Intel, outlined the commercial intent behind the architecture.
Inference is becoming the defining AI workload, and enterprises need infrastructure that delivers the right performance at the right cost. With Intel GPU D series, we are focused on making high capacity inference practical to deploy and economical to scale.
Intel pairs these accelerators with Xeon server designs built to the Open Compute Project Open Rack v3 specification. Specific Xeon 6+ rack setups scale to 46,000 cores while drawing 133.45 kW of total rack power. These CPUs manage system data routing, token orchestration, and background memory management to keep the accelerators fed with continuous calculations. Intel is also contributing an open adapter card specification to help cloud providers integrate third party accelerators into server racks without redesigning their physical power and cooling layouts.
Software deployment relies on the shared Xe software platform and standard industry frameworks. Developers receive early support for models like Meta Muse Glimmer alongside the Intel AI for Enterprise Agent Toolkit. This open framework allows code written for Xeon processors to execute across D series GPUs without full application rewrites. The approach lets data centers scale inference capacity while retaining existing x86 management infrastructure.
