Intel has officially named its next-generation AI accelerator the Intel GPU D-Series and released its first concrete performance specifications. The announcement came during the Open Computing Project (OCP) 2026 conference, where the company provided the first clear look at the capabilities of the chip, internally known as Crescent Island.
The new hardware targets a theoretical peak performance of 1.3 PFLOPS in FP8 precision. This metric measures the chip’s ability to perform floating-point operations, a standard benchmark for artificial intelligence workloads. Intel also confirmed that the accelerator will offer a memory bandwidth of 1.5 TB/s. These figures represent the final targets for the silicon, following a period of limited disclosure throughout 2026.
Crescent Island Architecture and Form Factor
The GPU D-Series is built on Intel’s Xe3P graphics architecture. Unlike traditional graphics processing units, this chip lacks many of the functional units required for rendering images. Instead, it is designed as a massively parallel computing accelerator optimized for data center tasks. The company plans to release the product in an air-cooled PCIe form factor with a thermal design power (TDP) of 350W.
Each unit will support up to 480GB of LPDDR5X memory. This configuration allows the chip to operate in standard rack servers without the complex liquid cooling infrastructure required by higher-power competitors. Intel states that these systems can scale up to eight GPUs per node, similar to existing high-density server architectures.
The performance envelope of the GPU D-Series differs significantly from rival products like Nvidia’s Vera Rubin. While Nvidia’s chips can reach 17.5 PFLOPS of FP8 throughput, they require kilowatt-level power draws and liquid cooling. Intel’s approach focuses on efficiency and compatibility with existing air-cooled data center environments.
Targeting Large Language Model Inference
Intel positions the GPU D-Series specifically for AI inference rather than training. The company claims the chip can handle inference for models with more than one trillion parameters when deployed in distributed configurations. This capability is critical for running large language models in production environments where latency and cost are primary concerns.
The hardware is designed to support long-context inference and concurrent services. Intel highlights its suitability for multi-tenant AI services, robotics, and other applications that require low-latency responses. The air-cooled design allows for easier deployment in facilities that may not have been upgraded for liquid cooling.
To support these workloads, Intel is leveraging its existing Xe software foundation. The company intends to use the same software stack across its integrated graphics and discrete GPU products to simplify development for independent software vendors and cloud providers. This strategy aims to reduce the friction of adopting new hardware in established data centers.
Intel is currently sampling the GPU D-Series to a broad ecosystem of original equipment manufacturers, original design manufacturers, and AI-native companies. The company did not provide specific details on launch dates or pricing. The success of this product will determine whether Intel can regain ground in the AI accelerator market after previous setbacks with its Data Center GPU Max family.
Source: Tom’s Hardware
