Intel has released new technical details of its D270 AI inference GPU, an air-cooled design intended for existing data center servers. Intel’s own card has 160 gigabytes of memory, while partner-built versions can offer configurations reaching 480GB. Customer sampling is underway.
The company presented the D-series, formerly code-named Crescent Island, at the Open Compute Project Global Summit on October 12. Its announcement advertises up to 480GB of LPDDR5X memory, with the lower capacity of the Intel-branded card specified in a footnote.
The design uses LPDDR5X rather than high-bandwidth memory, or HBM. Intel rates it at 1.5 terabytes per second of memory bandwidth and 1.3 petaflops of FP8 compute within a 350-watt power envelope.
D270 is built on Intel’s Xe3P architecture, with 32 compute cores and 256 matrix engines. Intel says it has reassigned chip area that would otherwise serve graphics and ray tracing to inference compute. The product is designed to run deployed models, including enterprise assistants and other applications that make repeated calls to models and tools.
A particular focus is prefill, the stage in which a model processes input before generating its answer. Long documents and accumulated context from multistep tasks increase that work. Intel also emphasizes memory capacity for serving multiple sessions and retaining the cached model state used during inference.
The card uses a standard PCIe 5.0 connection and passive air cooling. Intel says that allows evaluation and deployment in existing servers without a specialized liquid-cooling retrofit. The intended customers include enterprise infrastructure teams, AI service providers and sovereign-cloud operators.
Intel is working with server makers, contract hardware manufacturers, cloud providers and software vendors on validated platforms and integration. Its product page names PyTorch and vLLM among the software ecosystems it supports, while the common Xe software foundation is intended to help reuse applications across Intel GPUs.
The page’s comparison with Nvidia’s RTX Pro 6000 carries a separate qualification: performance projections are based on public information and Intel’s internal analysis as of February 2026, with results dependent on workload and configuration.
The announcement follows Microsoft’s Maia 300 accelerator unveiling. Microsoft plans multi-gigawatt Maia deployments starting in 2027; its new accelerator uses HBM4 and is organized into 96-chip racks.
Intel also presented Xeon-based rack designs at the summit. Select Xeon 6+ configurations exceed 46,000 CPU cores per rack at 133.45 kilowatts of total rack power. That figure describes separate CPU systems, not the power required by a D270 card.
–
By the Control Plane Editorial Team