Intel Names Crescent Island the GPU D270, Putting 480GB of LPDDR5X in Air-Cooled Racks

At the Open Compute Project Global Summit in San Jose on 12 October, Intel gave the accelerator previously known as Crescent Island a brand name, the Intel GPU D-Series, and published specifications for the first card in the family, the GPU D270. It is Intel's first commercial data-center GPU since Ponte Vecchio in 2021 and the first part built on Xe3P, an inference-tuned derivative of Xe3. The chip carries 32 Xe3P cores and 256 XMX matrix engines running at roughly 2.5GHz, with a 32MB unified L2 cache, a PCIe Gen5 x16 interface and a 350W rating in a passive air-cooled PCIe CEM card that drops into existing standard servers rather than requiring dedicated racks or liquid cooling.

The headline numbers are 1.3 PFLOPS of peak FP8 matrix compute and 1.5 TB/s of memory bandwidth, the latter delivered by a 20-chip, 80-channel LPDDR5X subsystem clocked at about 9,375MT/s across a 1,280-bit bus. Intel's own card ships with 160GB of LPDDR5X, while ODM partners can build versions with up to 480GB, and the part covers data formats from FP4 to FP64 as well as carrying a media block with four encoders and four decoders for multimodal pipelines. Xe3P redesigns the XMX engines with systolic arrays going four-deep to sixteen-deep, letting each engine do more multiply-accumulate work per operand fetch, and strips ray-tracing, 3D graphics and display hardware in favour of compute silicon.

The economics of the design rest on LPDDR5X rather than HBM. LPDDR5X costs less per gigabyte and draws less power but delivers less bandwidth, which explains both the 480GB capacity option and the 1.5 TB/s ceiling. Intel says a single card can serve models up to 300 billion parameters and an eight-GPU server models beyond one trillion, with sampling already underway to customers and no pricing or general availability announced. Independent analysis from Jon Peddie Research notes the arithmetic that model operators will need to check: a dense 300B FP8 model reading every weight per token tops out near five tokens per second per stream on 1.5 TB/s, so the claim rests on either the 480GB partner card or on FP4 weights squeezed onto the 160GB Intel board with little room left for KV cache.

Power and deployment are where Intel expects to win rather than raw throughput. At 350W the D270 sits far below the 600W of AMD's comparable PCIe inference card and below kilowatt-class liquid-cooled rivals, giving roughly 3.7 TFLOPS of FP8 per watt, a ratio that favours prefill and batched serving over single-stream decode. Intel also showed a rack design built to the OCP Open Rack Version 3 standard for agentic AI that accommodates more than 46,000 Xeon 6+ CPU cores at a 133.45 kW rating. Alongside the hardware the company is pitching cost rather than peak performance: AI infrastructure head Anil Nanduri said inference requires sufficient performance at an appropriate cost, with the D-series aimed at scaling large inference workloads cost-effectively.