Intel reveals GPU D-Series specs for air-cooled AI inference servers
Intel names the Crescent Island chip the GPU D-Series, offering 1.3 PFLOPS FP8 performance in a standard server form factor.
Key points
- Intel officially named the Crescent Island accelerator the GPU D-Series at OCP 2026.
- The chip delivers 1.3 PFLOPS of FP8 compute and 1.5 TB/s of memory bandwidth.
- Products will use air cooling and fit into standard PCIe slots with up to 480GB of memory.
Intel announced the official name and performance targets for its new AI accelerator at the Open Compute Project event in 2026. The company revealed that the chip previously known as Crescent Island will be marketed as the Intel GPU D-Series. This announcement provides the first confirmed theoretical peak performance numbers for the hardware, which is designed specifically for artificial intelligence inference workloads rather than general-purpose graphics rendering.
In plain English
The GPU D-Series is built on Intel’s new Xe3P graphics architecture, which the company describes as a massively parallel computing structure. Because the chip lacks most graphics-specific functional units, it operates more like a dedicated computing accelerator than a traditional graphics card. Intel states that the chip can deliver 1.3 petaflops of performance using FP8 precision, a data format commonly used for efficient AI processing.
The hardware includes up to 480 gigabytes of LPDDR5X memory, providing a total memory bandwidth of 1.5 terabytes per second. This large memory capacity allows the chip to handle large language models and long-context inference tasks without requiring excessive data transfers between the processor and external storage systems.
The background
Intel designed the GPU D-Series for deployment in traditional air-cooled rack servers, distinguishing it from high-performance chips that require liquid cooling systems. The accelerator features a thermal design power of 350 watts and uses a standard PCIe form factor. This design allows administrators to install up to eight GPUs in a single server chassis, similar to configurations used by competitors like Nvidia in their HGX systems.
According to Tom's Hardware, the GPU D-Series targets a different performance envelope than Nvidia’s Rubin GPUs, which offer significantly higher throughput but require complex liquid cooling infrastructure. Intel’s approach focuses on fitting AI capabilities into existing data center cooling and power constraints.
What changes now
Intel says the GPU D-Series can power inference on models with more than one trillion parameters when deployed in multi-GPU server configurations. The company highlights suitability for multi-tenant AI services, robotics applications, and other latency-sensitive workloads that require concurrent inference services. Customer sampling of the GPU D-Series products is currently in progress, although Intel has not announced a specific date for a broad commercial launch.
To support these deployments, Intel is leveraging its common Xe software foundation, which it has developed for AI tasks across its integrated and discrete GPU products. The company states it is working with a broad ecosystem of original equipment manufacturers, cloud providers, and independent software vendors to integrate the accelerators into enterprise environments.
What to do and how to stay safe: Intel
- Review your data center cooling capacity to determine if air-cooled 350W accelerators fit within current thermal limits for rack servers.
- Evaluate whether your AI workloads require the large memory capacity of 480GB to handle long-context models without external storage bottlenecks.
- Monitor vendor announcements for software stack compatibility, ensuring your existing AI frameworks support Intel’s Xe3P architecture and common Xe foundation.
- Plan for PCIe slot availability in your server chassis, as the GPU D-Series requires standard PCIe form factor installation rather than proprietary rack-scale interconnects.
Step-by-step guide: Stop AI Chatbots Leaking Sensitive Business Data
General security guidance from the Vector Update newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is the official name for Intel's Crescent Island chip?
Intel officially named the Crescent Island accelerator the GPU D-Series at the OCP 2026 event.
How much memory does the GPU D-Series support?
The GPU D-Series supports up to 480GB of LPDDR5X memory with 1.5 TB/s of bandwidth.
Does the GPU D-Series require liquid cooling?
No, the GPU D-Series is designed for air-cooled PCIe form factors with a TDP of 350W.
