Intel D270 data-center GPU gets performance and memory specifications

Intel has named the Crescent Island GPU family the Intel GPU D-Series, and the D270 is the product formerly known as Crescent Island.[2][3] Intel describes the D270 as an inference-optimized data-center GPU.[4] Tom’s Hardware reported that the chip’s performance and bandwidth specifications were presented at OCP 2026.[1] Intel says the D270 is designed for agentic AI, long contexts and high-concurrency serving.[4] These descriptions position the product around inference workloads, rather than establishing performance across every type of AI task.[4]

Tom’s Hardware reported peak FP8 compute of 1.3 PFLOPS and memory bandwidth of 1.5 TB/s.[1] TechPowerUp reported support for up to 480 GB of LPDDR5X memory, 32 Xe3P cores, 256 XMX engines and 32 MB of unified L2 cache.[2] The figures describe different parts of the design: compute throughput, bandwidth, memory capacity and chip resources.[1][2] Teams comparing accelerators can keep those measures distinct and evaluate them against their own workload requirements.

Particle characterized the D270 design as air-cooled, using LPDDR5X memory and PCIe Gen5.[3] It also reported that the design prioritizes memory capacity over HBM-level bandwidth, and could perform less well on workloads that are hungry for bandwidth.[3] That qualification matters alongside the headline compute and memory figures: the reported peak FP8 compute does not, by itself, establish performance for a particular application.[1][3] A workload assessment should therefore consider whether its memory-capacity needs outweigh its sensitivity to bandwidth.

Particle reported that Intel had begun customer sampling of the D270.[3] The evidence does not specify general availability, a price or a customer ship date. Intel says the D-Series, Xeon CPUs, open software and ecosystem innovations are intended to help customers scale agentic AI.[5] That describes Intel’s broader positioning, not a reported benchmark or proof of results from a customer deployment.[5] For teams considering an evaluation, sampling is the availability detail established here; other purchasing and deployment details remain unreported in the evidence.