Intel Unveils GPU D-Series AI Inference Accelerator With 480GB Memory in Air-Cooled PCIe Design

The physical limit of artificial intelligence deployment is no longer computational speed, but thermal dissipation and memory capacity inside the server rack. When an enterprise attempts to run a frontier language model, the system must hold hundreds of billions of parameters in active memory simultaneously to function.

If those parameters exceed the memory capacity of a single accelerator, the model must be split across multiple chips, multiplying the hardware cost and power draw. Memory capacity dictates whether a model can run efficiently or if it becomes bottlenecked by constant data transfers between separate nodes.

Intel has targeted this exact deployment bottleneck with the introduction of its GPU D-Series, previously code-named Crescent Island. Announced at the 2026 Open Compute Project Summit, the new D270 accelerator abandons traditional graphics processing entirely to focus exclusively on enterprise AI inference workloads.[2][7]

The defining specification of the D270 is its massive memory payload: up to 480 gigabytes of LPDDR5X memory per card. This represents a significant departure from standard high-bandwidth memory configurations, deliberately trading peak data transfer speeds for sheer volume and capacity.[3][4]

By utilizing LPDDR5X—the same power-efficient memory standard used in high-end laptops—Intel achieved a memory bandwidth of 1.5 terabytes per second. While slower than the HBM3e memory found on Nvidia's flagship chips, the 480-gigabyte capacity allows a single card to hold massive models entirely in local memory.[1][2]

The Air-Cooled Advantage

The most consequential design choice for enterprise adoption is the card's physical format. The GPU D-Series is built as a standard PCI Express (PCIe) add-in card, identical in shape and installation method to a conventional high-end graphics card used in standard workstations.[2][5]

More importantly, the hardware is entirely air-cooled and operates within a strict 350-watt power envelope. The current generation of frontier AI accelerators, such as Nvidia's Blackwell Ultra or AMD's MI455X, draw so much power that they increasingly require direct-to-chip liquid cooling infrastructure to prevent overheating.[2][6]

Liquid cooling demands specialized data center plumbing, reinforced floor loading, and custom server racks that most traditional enterprises do not possess. By capping the power draw and utilizing standard fans, the D270 can be slotted directly into existing corporate server racks without facility upgrades.[2][7]

"Inference is becoming the defining AI workload, and enterprises need infrastructure that delivers the right performance at the right cost," said Anil Nanduri, vice president of AI Products at Intel's Data Center Group, during the product's official unveiling at the summit.[7]

Nanduri noted that the company designed the hardware specifically to integrate into standard enterprise environments. "With Intel GPU D-series, we're focused on making high-capacity inference practical to deploy and economical to scale," he stated, emphasizing the avoidance of custom cooling loops.[7]

Targeting the 300-Billion Parameter Class

The computational core of the D270 delivers 1.3 peak petaflops of FP8 compute performance. This specific level of processing power is calibrated for inference—the act of generating responses from an already-trained model—rather than the highly intensive initial training phase that requires massive supercomputers.[1][2]

The 480-gigabyte memory buffer was not chosen arbitrarily by Intel's engineering team. It is precisely sized to accommodate the current generation of open-weight enterprise models, particularly those operating in the highly capable 70-billion to 300-billion parameter range.[4][7]

Models like Meta's Llama 3 require roughly one byte of memory per parameter when quantized to 8-bit precision. A 300-billion parameter model therefore needs over 300 gigabytes of active memory just to load, before accounting for the context window of the user's prompt and generated output.[4][6]

Previously, running a model of that size required networking four or eight standard 80-gigabyte accelerators together. The D270 allows an enterprise to run these massive models on a single node, drastically reducing the total cost of ownership and the complexity of the underlying software stack.[2][7]

The architecture relies on Intel's Xe3P graphics architecture, but the silicon has been stripped of all display outputs, rasterization engines, and 3D rendering hardware. Every millimeter of the die is dedicated exclusively to the matrix multiplication and tensor operations required by modern neural networks.[3][6]

Open Standards and Market Position

Intel's strategy leans heavily on open industry standards to counter the proprietary ecosystems of its rivals. The D-Series utilizes the Open Compute Project's specifications where applicable, though the primary enterprise push relies on the universal PCIe standard that every major server manufacturer already supports.[2][7]

The hardware also supports the open-source SYCL software stack, allowing developers to write code that can run across different types of accelerators. This open approach prevents enterprises from being locked into a single vendor's proprietary programming language when deploying their internal AI tools.[7]

Market analysts note that Intel is not attempting to beat Nvidia's flagship chips in absolute benchmark performance. Instead, the company is targeting the massive middle market of enterprise customers who need to deploy AI locally for privacy, regulatory compliance, and data security reasons.[5]

"Agentic AI is changing what customers need from their data centers," said Kevork Kechichian, executive vice president of Intel's Data Center Group. "Success will depend not on any single component, but on open, heterogeneous infrastructure that brings together Xeon, purpose-built acceleration and interoperable software."[7]

The ultimate success of the Crescent Island architecture will depend heavily on its final pricing and volume availability. While the specifications offer a clear path for enterprise deployment, the hardware must still compete with aggressive pricing from cloud providers offering inference as a remote service.[1][5]