Intel D270 GPU AI Accelerator–Up To 480 GB LPDDR5X

TL;DR

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Intel is presenting its D270 as an inference-focused data-center GPU, with up to 480GB of LPDDR5X memory, 350-watt air-cooled operation and a standard PCIe 5.0 form factor. The company says the design targets long-context and high-concurrency AI workloads; independent performance results, pricing and availability are not provided in the supplied material.

Intel is promoting its D270 as a data-center GPU for production AI inference, specifying up to 480GB of LPDDR5X memory, 350-watt air-cooled operation and a standard PCIe 5.0 form factor. The company says the design is intended for long-context and high-concurrency workloads, including agentic AI, but its published material does not provide a release date, price or independent performance results.

Intel describes the D270 as an inference-optimized accelerator rather than a product aimed primarily at training large models. Its listed compute configuration includes 32 Xe3P cores and 256 XMX matrix engines. Intel says the architecture reallocates die area from graphics and ray-tracing pipelines toward compute and matrix operations, with a focus on prompt processing, or prefill, and serving multiple sessions. The source does not include benchmark results to quantify those claims.

The headline memory figure is up to 480GB of LPDDR5X per card. Intel argues that high capacity can help keep more model data and long-context workloads resident on a GPU, potentially reducing the number of accelerators needed for some deployments. The company says the card does not use high-bandwidth memory (HBM); the supplied product description gives no memory-bandwidth figure, capacity breakdown by configuration or independent comparison of performance.

Intel also lists a 350W power rating, passive air cooling and standard PCIe 5.0 connectivity. It says the design is intended to fit existing data-center servers without proprietary networking or a liquid-cooling retrofit. Intel identifies PyTorch and vLLM among the software ecosystem relevant to deployment, but the supplied material does not provide a compatibility matrix, availability details or results from customer deployments.

Memory Capacity Shapes Inference Plans

The D270’s pitch addresses a practical constraint for production AI: inference systems must hold model weights, prompts and intermediate data in memory while serving users. Long prompts and concurrent requests can increase memory demand, particularly in retrieval-heavy applications and agent workflows that make repeated model calls. Intel’s emphasis on capacity suggests it is targeting deployments where fitting more workloads on each card could matter as much as peak compute speed.

That positioning may interest enterprise, cloud and government operators that need to expand inference capacity within existing rack power and cooling limits. A standard PCIe card with air cooling could be easier to trial in conventional server environments than equipment that requires specialized infrastructure. Those are potential deployment advantages, not demonstrated savings: whether the D270 delivers more tokens per watt or lower cost per request will depend on model, software, configuration and measured workload performance.

The product also gives buyers another architecture to evaluate as demand shifts toward serving deployed models, rather than only training them. The relevant comparison is not memory capacity alone. Memory bandwidth, software maturity, throughput, latency, power use and price will determine whether the card is competitive for a particular service.

Intel’s Focus on Agentic Inference

Intel frames the D270 around agentic AI, in which a system may reason over a long prompt, retrieve information, use tools and make several model calls before answering. Such workflows can create substantial prefill work and keep context or cached data in memory across a session. Intel says the D270 is designed for these demands, as well as retrieval-augmented generation, assistants, document understanding, multimodal applications and diffusion models.

The company presents the D270 as part of its D-series data-center GPU line and emphasizes standard server integration: PCIe 5.0, air cooling and open software frameworks. These details describe the intended product profile, not proof that every listed application or framework is supported in a particular configuration. Intel’s source also includes a performance-projection disclaimer referring to comparisons with the RTX Pro 6000 based on public information and Intel internal analysis as of February 2026; it says results vary by use and configuration and specifies a memory-capacity comparison at comparable bandwidth. No numerical projection is included in the material provided here.

““Intel GPU D270 is an inference-optimized data center GPU designed for the agentic AI era.””

— Intel

Pricing, Benchmarks and Availability

The product information supplied does not state when the D270 will ship, which configurations will be sold, how much they will cost or which markets and system partners will offer it. Intel has not provided, in this material, independent benchmarks for throughput, latency, tokens per watt or performance across specific model sizes. The listed capacity is “up to” 480GB, and the source does not clarify which configurations reach that maximum.

It is also unclear how the card’s LPDDR5X bandwidth and software support compare with competing accelerators on real workloads. Intel’s product claims describe design goals; they should not be treated as verified results. Buyers will need workload-specific testing to assess model compatibility, concurrency, reliability and total cost.

Awaiting Sales and Test Details

Intel says prospective customers can contact the company for more information and updates as product details become available. The next useful milestones would be confirmed availability, pricing and system configurations, followed by public benchmark data and independent testing on representative inference workloads. Those details will show whether the D270’s memory capacity and air-cooled profile translate into practical gains for data-center operators.

For now, the announcement establishes Intel’s stated specifications and target use cases, but not a delivery schedule or measured value proposition. Deployment decisions remain premature until buyers can verify software support and compare performance, power and cost against alternatives using their own models and serving patterns.

Key Questions

What is the Intel D270?

Intel describes it as an inference-focused data-center GPU for production AI applications, including long-context and high-concurrency serving. The supplied material does not present it as a GPU primarily designed for large-scale model training.

How much memory does the D270 have?

Intel lists up to 480GB of LPDDR5X per card. The source does not specify which configurations provide that maximum or give a memory-bandwidth figure.

Does it require liquid cooling?

Intel lists the D270 as a 350W, air-cooled card and says it is designed for standard PCIe 5.0 infrastructure. Actual server compatibility will depend on system configuration.

When will the D270 be available, and what will it cost?

The supplied Intel material does not state a launch date or price. The company says customers can contact Intel for further information and product updates.

Are Intel’s performance claims independently verified?

The provided material describes Intel’s product positioning and intended workloads, but does not include independent benchmark results. Performance and cost claims need testing on specific models, software and serving configurations.

Source: hn