Intel D270 AI Accelerator Debuts With Up to 480GB Memory for Large Language Model Inference
Intel has introduced the D270, the first accelerator in its new GPU D-series, a data-center product designed specifically for AI inference. Previously known under the codename Crescent Island, the D270 is not aimed at gaming or traditional graphics rendering. Instead, it is built to handle demanding artificial intelligence workloads, especially large language models, long-context processing, and high-volume AI services.
The most eye-catching feature of the Intel D270 is its memory capacity. Intel’s reference design includes 160GB of LPDDR5X memory, while custom versions from hardware partners can scale up to an enormous 480GB. That amount of memory could make the card particularly attractive for companies running large AI models that need to keep massive model weights readily available while also supporting extended conversations and complex prompts.
Unlike many high-end AI accelerators that rely on HBM memory, Intel has chosen LPDDR5X for the D270. This decision appears to focus on balancing capacity, cost, and deployment flexibility. For inference workloads, especially those involving large language models, having more memory can be just as important as raw bandwidth. Intel is positioning the D270 as a practical option for enterprises and cloud providers that need scalable AI performance without moving to more complex server designs.
The D270 delivers up to 1.5TB/s of memory bandwidth and offers peak FP8 performance of 1.3 petaflops. It operates within a 350W power target and uses a PCIe 5.0 x16 form factor with air cooling. That is an important detail for data centers, because it means the accelerator can fit into many existing server environments without requiring specialized liquid-cooling infrastructure.
Under the hood, the accelerator is based on Intel’s Xe3P architecture. It includes 32 Xe3P cores, 256 XMX engines for matrix acceleration, and 32MB of unified L2 cache. Intel has removed traditional 3D graphics and ray-tracing hardware from this design, freeing up more chip area for compute-focused AI processing. As a result, the D270 is clearly not a consumer graphics card, even though it is built using GPU technology.
Intel says the GPU D-series is designed for a wide range of AI inference tasks, including enterprise assistants, coding assistants, document analysis, multimodal AI, diffusion models, agentic AI systems, and high-concurrency AI services. A single D270 card is said to support models with up to around 300 billion parameters, while an eight-card server configuration could handle models exceeding one trillion parameters.
This focus on inference is important. While AI training often gets the spotlight, inference is where models are used in real-world applications. Every chatbot response, AI coding suggestion, document summary, image generation request, or enterprise automation task depends on inference performance. As businesses adopt AI tools at scale, hardware that can serve more users efficiently is becoming increasingly valuable.
Intel has confirmed that customer sampling is already underway with server manufacturers, cloud providers, and software companies. However, the company has not yet announced official pricing or a broad release date for the GPU D-series.
With the D270, Intel is taking a targeted approach to the fast-growing AI accelerator market. Rather than competing purely on graphics features or training performance, the company is focusing on high-capacity memory, air-cooled deployment, and large-scale AI inference. If Intel can deliver competitive pricing and strong software support, the D270 could become a serious option for data centers looking to run massive AI models more efficiently.
