Intel launches D270 data center GPU with up to 480 GB LPDDR5X for AI inference

Intel targets inference with a memory-heavy new GPU

Intel has published details of the GPU D270, a data center accelerator designed specifically for running AI models in production rather than training them at scale. The product page, which surfaced on the front page of Hacker News, lists a headline specification that stands out in the current accelerator market: up to 480 GB of LPDDR5X memory, rather than the high bandwidth memory (HBM) used in most AI GPUs. The card runs at 350 W with passive air cooling on a standard PCIe 5.0 interface.

A memory-capacity-first design

Intel's central argument is that agentic AI has changed the profile of data center workloads. Because agents consume far more input than they generate — long prompts, retrieval-heavy context, many sequential model calls — the compute burden shifts toward prefill processing, and the memory burden shifts toward holding large models, long contexts and growing KV caches. According to Intel, the D270 is built for exactly this pattern: memory-rich, power-efficient serving where keeping more of the model and its context resident matters more than raw training throughput.

The choice of LPDDR5X over HBM reflects that positioning. Intel says the D270 can keep whole-model and long-context workloads resident across fewer GPUs with no HBM at all. The trade-off is bandwidth: LPDDR5X generally delivers substantially lower memory bandwidth than the HBM found in competing accelerators, and Intel's page does not publish bandwidth figures, so real-world throughput on bandwidth-hungry token generation remains an open question until independent benchmarks appear.

Architecture and precision

The D-series silicon is derived from Intel's Xe graphics lineage, but with die area reallocated away from graphics and ray-tracing pipelines toward general Xe3P cores and matrix compute. The D270 carries 32 Xe3P cores and 256 XMX engines, which Intel pitches as delivering more prompt processing per watt for long-context and agentic workloads. Precision support spans native low-precision inference up through FP64 for scientific computing, with Intel framing the workload, not the hardware, as the thing that should determine which format is used.

Standard racks, familiar software

Deployment simplicity is a repeated theme on the product page. The card uses a standard form factor and standard PCIe 5.0 interconnect in existing servers, with no proprietary networking or bespoke systems, and no liquid-cooling retrofit between evaluation and production. That matters for buyers who want to add inference capacity within current power and space budgets rather than building new facilities.

On the software side, Intel claims zero-day support for leading open models and points to the frameworks teams already run, including vLLM, PyTorch, SGLang and OpenAI Triton. The listed target workloads include agentic AI, long-context retrieval-augmented generation, enterprise and coding assistants, document understanding, multimodal AI, diffusion models and high-concurrency serving. Target customers range from enterprise AI and SaaS providers to GPU-as-a-service operators and government or sovereign-cloud deployments.

What Intel has not disclosed

The launch material omits several things buyers will want before committing: pricing, availability dates, and performance or efficiency numbers against competing inference hardware. The page is also light on interconnect details for multi-GPU scale-out, which matters for serving very large models. Until third-party benchmarks exist, the claim that LPDDR5X-based inference can be both memory-rich and fast enough for high-concurrency serving is Intel's positioning rather than a demonstrated result.

Why it matters

Inference, not training, is where AI infrastructure costs are accumulating, and agentic systems multiply that cost by turning one user request into many rounds of model calls, retrieval and reasoning. HBM supply is expensive and constrained, and it sets a hard ceiling on memory capacity per accelerator. A design that pairs 480 GB of capacity with a 350 W, air-cooled, standard-PCIe envelope attacks the economics of long-context and high-concurrency serving from a different angle than the HBM-first incumbents — keeping agents resident, raising tokens per rack and per watt, and letting enterprises use the data centers they already run.

The D270 is also Intel's latest attempt to carve out a credible position in data center AI acceleration. Whether LPDDR5X's bandwidth is sufficient for production token serving at competitive cost is the question the market will now answer.