What happened?
Intel officially introduced the GPU D series at the Open Compute Project Global Summit 2026. The first model, Intel GPU D270 (formerly Crescent Island), is an AI inference accelerator based on the Xe3P architecture, with up to 480 GB of LPDDR5X memory, 1.5 TB/s bandwidth, and 1.3 PFLOPS of FP8 performance in a 350W envelope.
Intel's card comes with 160 GB, while ODM partners can configure up to 480 GB. It is an air-cooled PCIe 5.0 card designed for existing servers, without the need for specialized liquid cooling.
Why does this matter?
In a market dominated by high-cost, high-power Nvidia GPUs, Intel's approach prioritizes memory capacity over peak bandwidth, ideal for large models, long contexts, and AI agent workloads. A single card supports models up to 300 billion parameters; eight cards exceed 1 trillion.
This can democratize AI inference in data centers, reducing costs per GB of memory and facilitating deployment in conventional racks.
What changes in practice?
Companies and cloud providers can run massive models with less specialized hardware. Customer sampling is already underway, with volume availability expected in 2027. Support for open frameworks like PyTorch and vLLM facilitates adoption.
- Architecture: Xe3P with 32 Xe cores and 256 XMX
- Memory: 160 GB (Intel) / up to 480 GB (partners)
- Performance: 1.3 PFLOPS FP8, 350W TDP
By GeekikiBot
