Skip to content

GPU Platforms

Intel D-Series at OCP 2026: Read the Card Behind the 480 GB Headline

Intel's OCP release distinguishes its 160 GB card from partner designs up to 480 GB. What sampling and PCIe positioning let enterprise buyers evaluate.

Leviathan SystemsPublished 2026-10-205 min read

Analysis checked October 20, 2026.

Intel's October 12 OCP announcement gives enterprises new details on D-series, previously code-named Crescent Island, with a consequential qualification: Intel says its own branded PCIe card has 160 GB of LPDDR5X, while the design enables partner ODM-branded cards with options up to 480 GB. The largest figure describes a design possibility; it does not identify every card a buyer might receive. [1, 3]

Our assessment: D-series makes a memory-rich accelerator in conventional server infrastructure a concrete evaluation topic. The purchase still has to identify an exact board inside a supported complete server. That transition matters especially to enterprises hoping to reuse an existing server estate: a family announcement can start the investigation, while the supplied configuration determines what can actually be assessed.

Attach the capacity to the offered board

The qualification changes the first supplier question. Ask which manufacturer's board and memory configuration the response covers, then retain that identity throughout the server and application evaluation. A proposal using the Intel-branded configuration and a proposal using a partner's higher-capacity design describe different hardware. Neither the family name nor an image of a card reconciles them.

Intel's current D270 page identifies LPDDR5X and advertises up to 480 GB per card. [2] That makes the local memory budget an evaluation question: can the offered configuration keep the required model and live inference state resident? LPDDR5X and HBM are different memory technologies; capacity and the rate of data movement remain separate specifications. More capacity alone cannot establish the application's latency or throughput.

For comparisons with other accelerators, the existing selected GPU memory specification guide keeps named variants, memory types and system boundaries explicit. Use that comparison for its documented selection. This announcement adds a new product question to investigate; it supplies no basis for inserting the family ceiling as the specification of an unidentified offered card.

A conventional card shifts the server question

Intel positions D270 around PCIe 5.0 and a 350 W, passively air-cooled design. [2] For an enterprise with conventional air-cooled servers, that is a useful starting point: the proposed route centers on a card and host combination rather than beginning with a specialized liquid-cooled rack. It gives the infrastructure team a particular configuration to ask an OEM about.

The physical mechanism still runs through the host. A passive card's heat must be carried away by the surrounding cooling arrangement. An available expansion connector alone says nothing about whether the proposed server provides the required space, power connections and airflow. Request the OEM's supported configuration for the actual board, GPU count and host revision. The air-cooled description also leaves the room's capacity to support that complete system to be established.

Reuse therefore needs a configuration-specific response. The configured-server versus GPU upgrade-kit guide explains why a board purchase can leave internal enablement hardware and product work unresolved. For D-series, settle whether the supplier proposes a complete configured server or a supported change to an identified owned host before estimating the benefit of keeping the existing chassis.

Sampling, system support and supply answer different questions

In its October 12 release, Intel says customer sampling is underway. [1] As reviewed on October 20, that is an attributed statement about evaluation activity. It does not tell a particular enterprise whether it can obtain a sample, which card configuration would be supplied or what access conditions apply.

An evaluation opportunity also differs from an OEM-supported production configuration. The latter needs an identified complete server, the supplied board population and the applicable support basis. The application owner then needs the software configuration and evidence relevant to the intended use. A sample can help answer an application question without establishing that the same hardware and software combination is an offered production system.

A supplier's delivery response answers another question: what equipment, quantity and delivery conditions will that supplier actually commit to? Public sampling language contains no project allocation. Conversely, a proposed shipment date cannot resolve missing server-support evidence. Keep those answers attached to the same configuration so that a promising evaluation does not silently become a purchase of different hardware.

Two replies to the same inquiry

Consider a fictional enterprise asking whether D-series should advance to application evaluation for its existing air-cooled server estate. It requests the actual supplied memory configuration, a supported complete server and the terms of evaluation access. The following replies are invented to illustrate the comparison; they contain no real quotation, inventory or test result.

Supplier A names a complete supported server model in its response, identifies the board manufacturer and exact supplied memory configuration, states the GPU count, and provides the applicable OEM configuration document. It separately describes whether the proposed evaluation equipment is a sample and what would still require confirmation for production supply.

Supplier B repeats the family's maximum memory capacity and its conventional-card positioning. It leaves the supplied board and host unidentified, and treats the announcement's sampling statement as its answer about availability.

A gives the buyer something assessable: the documented system can be checked against the requested application evaluation, and the remaining supply question is visible. B leaves the hardware comparison open. The buyer cannot credit B with the maximum capacity as an included specification, compare the two complete systems or schedule installation around its response.

The useful follow-up to B is the missing product identity and support evidence. B might eventually propose a suitable system; A might still fail the application's requirements. The difference at this stage is that A defines what the next evaluation would cover. If A's response concerns a newly supplied server, it also leaves the buyer's original reuse objective unresolved until the owned-host alternative is addressed.

Advance the evaluation before planning installation

The October announcement is most useful as a reason to request a specific evaluation package. Ask which board and complete host can be made available, which software configuration accompanies them, and which application requirement the evaluation should answer. Keep the workload owner's capacity, response-time and output-quality requirements with that request. A memory-rich design earns consideration when it addresses a defined constraint; the announcement alone supplies no measured application outcome or economic comparison.

After actual equipment selection, Leviathan's rack-and-stack service provides a route to discuss mounting, agreed physical IT cabling, labels and installation records for customer-furnished equipment. Bring the selected complete configuration, appointed equipment supplier, rack layout, cable endpoints and site window to the project inquiry. That makes the physical scope reviewable against equipment the project has actually identified.

Contact

Discuss Your Infrastructure Project

Tell us about your project, and our team will follow up to discuss the scope, requirements and availability.