Intel has unveiled its GPU D-series accelerators, led by the D270, promising substantial memory capacity and AI inference performance within a 350W power envelope.
Previously known as Crescent Island, the family uses Xe3P architecture and targets data centres running trained AI models. Chipzilla disclosed the latest specifications at the Open Compute Project Global Summit on 12 October 2026.
Intel claims peak FP8 computing performance of 1.3 petaflops and memory bandwidth of 1.5TB/s, though those figures don’t reflect real-world performance. Comparisons with competing cards require equivalent precision, sparsity settings and software, which marketing tables occasionally treat as optional extras.
The design supports up to 480GB of LPDDR5X memory, giving larger models and their working data considerable room to breathe. Intel’s own branded PCIe card carries 160GB, while partners can build configurations that reach the advertised maximum.
Chipzilla chose LPDDR5X for its balance of capacity, power consumption, and cost, avoiding the expensive high-bandwidth memory used by many competing accelerators. That choice introduces a bandwidth trade-off, making workload a key consideration when assessing any promised savings.
The GPU contains 32 Xe3P cores organised into four compute slices, with eight cores within each slice. Together, these cores provide 256 vector engines and 256 Xe Matrix Extensions engines for general computing and matrix operations.
Its memory hierarchy includes 512KB of L1 cache and shared local memory per Xe core, providing 16MB across the chip. A unified 32MB L2 cache connects the compute slices, and it supports numerical formats from FP4 through FP64.
Chipzilla has removed graphics and ray-tracing pipelines, reallocating silicon to computing resources suited to the accelerator’s intended AI workloads. The standard PCIe 5.0 card uses air cooling, allowing deployment in suitable existing servers without requiring a liquid-cooling conversion.
The architecture emphasises prefill, the stage where a model processes an incoming prompt before generating a response. Large memory capacity helps accommodate lengthy contexts and the key-value cache used to retain information during inference.
Software support includes vLLM, SGLang, and llm-d, along with orchestration across different processors and accelerators using Nvidia Dynamo. Intel’s common Xe software foundation is intended to help developers reuse applications across its GPU platforms.
Intel Data Centre Group vice president for AI products and go-to-market Anil Nanduri said: “Inference is becoming the defining AI workload, and enterprises need infrastructure that delivers the right performance at the right cost.” Nanduri added: “With Intel GPU D-series, we’re focused on making high-capacity inference practical to deploy and economical to scale.”
Customer sampling is already underway, with Chipzilla working with server manufacturers, cloud providers and software companies on integration. Intel says supported configurations can accommodate models with up to 300 billion parameters on one card and more than one trillion across eight GPUs.
