Intel launches GPU D270 inference card with up to 480GB memory at OCP Global Summit

Intel formally launched the GPU D270 at the OCP Global Summit on October 12, targeting inference-heavy data center workloads with a memory-first design that undercuts rivals on cost. The card, previously codenamed Crescent Island, arrives after Intel canceled two earlier GPU projects, marking a cautious re-entry into a market dominated by Nvidia and AMD.

The D270 uses Intel's Xe3P architecture, built specifically for inference rather than training. It ships with up to 480GB of LPDDR5X memory on ODM partner variants, paired with 1.5 TB/s of memory bandwidth and 256 XMX engines. Peak compute lands at 1.3 PFLOPS of FP8 matrix performance. Intel's own reference build starts at 160GB, with the higher-capacity models coming through third-party manufacturers. Customer sampling is already underway, and broader availability is targeted for 2027.

A memory play for large language model serving

The 480GB configuration directly addresses the bottleneck that frustrates many inference teams: fitting large models into a single card. Hosting a model with hundreds of billions of parameters typically requires splitting it across multiple GPUs, adding latency and complexity. The D270's memory capacity lets operators keep those models on one accelerator. The 1.5 TB/s bandwidth sits below what HBM-based solutions offer, but remains sufficient for most inference pipelines. Intel is betting that memory density and cost will matter more to buyers than raw throughput.

Air cooling and standard PCIe integration

The card operates at 350W with air cooling in a standard PCIe 5.0 form factor. That removes the need for liquid cooling infrastructure, which often slows adoption of high-end AI hardware in existing data centers. Rack integration becomes a straightforward swap rather than a facilities project. The initial 160GB version targets mainstream inference deployments, while the 480GB ODM variants aim at hyperscalers and enterprises with extreme memory requirements.

Positioning against Blackwell Ultra and MI455X

Intel does not try to match Nvidia's Blackwell Ultra or AMD's MI455X on peak compute. The D270 sacrifices teraflops for a lower price point and denser memory, a trade-off that fits inference-specific workloads. The chip's cost efficiency could appeal to teams running high-volume LLM serving where per-query economics matter more than training speed. Intel's OpenVINO toolkit and oneAPI ecosystem provide the software layer, with the 2027 production timeline allowing room to mature the stack before volume shipments.

Why this matters for IT and operations teams

For operations groups managing inference fleets, the D270 introduces a PCIe air-cooled option that sidesteps the cooling and power headaches of higher-wattage accelerators. The 480GB memory ceiling means fewer cards per server for large-model serving, which simplifies procurement, reduces node count, and cuts interconnect complexity. If Intel hits its cost targets, the card could shift the math on inference hardware decisions - especially for teams that have been forced into multi-GPU setups purely for memory headroom. Early performance benchmarks will determine whether the architecture delivers on its throughput promises, but the integration story alone makes it worth tracking as sampling continues.