Intel has introduced the GPU D270, an inference-optimized data center accelerator designed for the agentic AI era. According to the product documentation, the GPU offers up to 480 GB of LPDDR5X memory per card, a specification intended to meet the high memory demands of agentic AI workloads where long contexts and numerous concurrent sessions require substantial memory retention to sustain throughput. The FAQ section confirms the GPU is designed for production AI inference workloads such as agentic AI, long-context retrieval-augmented generation (RAG), enterprise AI assistants, coding assistants, document understanding, multimodal AI, diffusion models, and high-concurrency serving. This positions the D270 as a solution for memory-bound inference scenarios in data center environments.
The image shows the physical D270 GPU, which features the 480 GB memory capacity described in the article as key for handling large agentic AI contexts.
The D270's architecture features 32 Xe3P cores and 256 XMX matrix engines, which Intel states are acceleration units for prompt processing (prefill) in agentic AI workflows. The company explains that agentic AI often involves agents reading far more than they write, making prefill compute dominant, and that die area has been reallocated from graphics and ray-tracing pipelines to these cores and engines to improve inference efficiency. This focus on compute aims to deliver more prompt processing per watt, curbing long-context and agentic AI costs, and the GPU supports a range of precision levels from native low-precision inference to FP64 scientific computing as the workload determines. Intel's documentation states the GPU is 'Built for compute' and 'Built for the workload' to emphasize this design philosophy.
This image illustrates the chip architecture mentioned in the article, where the die area is allocated to cores and engines for efficient agentic AI prefill compute.
Deployment considerations include a standard PCIe 5.0 form factor and 350W air-cooled operation, allowing the GPU to be integrated into existing data center servers without requiring liquid-cooling retrofits or specialized infrastructure. Intel also notes that the D270 supports open AI frameworks such as PyTorch and vLLM (with zero-day support for leading open models), and highlights the GPU's built-for-efficiency design that uses standard servers and avoids proprietary networking or bespoke systems. The product further emphasizes 'Built for your data center' and 'Open by design' to underscore its compatibility with existing enterprise infrastructure. However, the product page does not disclose pricing information or provide detailed performance benchmarks comparing the GPU to competing solutions.