How does Intel's air-cooled GPU D-Series with 1.3 PFLOPS aid enterprise AI?
Intel unveils the GPU D-Series, an air-cooled AI accelerator delivering 1.3 PFLOPS FP8 compute and 480GB memory for enterprise data centers.
Key points
- Intel officially names its Crescent Island AI accelerator the GPU D-Series at OCP 2026.
- The chip offers 1.3 PFLOPS of FP8 performance and up to 480GB of LPDDR5X memory.
- Designed for air-cooled servers, it targets multi-tenant AI services and large language models.
Intel has formally introduced the GPU D-Series, branding its previously code-named Crescent Island AI inference accelerator for the commercial market. According to Tom’s Hardware, the announcement occurred at the Open Compute Project 2026 event, where the company revealed specific performance metrics and deployment targets for the new hardware family.
What are the performance specifications?
The GPU D-Series is rated for a theoretical peak performance of 1.3 PFLOPS in FP8 precision. Tom’s Hardware reports that the chips also feature up to 480 GB of LPDDR5X memory, delivering 1.5 TB/s of memory bandwidth. This configuration supports the high-throughput requirements of modern artificial intelligence workloads while maintaining a power envelope suitable for standard data center infrastructure.
How does it compare to competitors?
Unlike Nvidia’s Rubin GPUs, which require liquid cooling and deliver 17.5 PFLOPS per chip in rack-scale systems, the Intel GPU D-Series is designed for traditional air-cooled rack servers. Tom’s Hardware notes that these accelerators scale up to eight GPUs per server, similar to Nvidia’s HGX systems. This design choice targets a different performance tier, prioritizing ease of deployment in existing facilities over maximum raw compute density.
What workloads is it intended for?
Intel states that the GPU D-Series is optimized for large language models, long-context inference, and concurrent inference services. According to Tom’s Hardware, the hardware aims to support multi-tenant AI services, robotics, and other latency-sensitive applications. The company also indicates that configurations using multiple GPUs can power inference on models exceeding one trillion parameters, enabling distributed processing across standard server racks.
What is the development status?
Customer sampling for the GPU D-Series is currently in progress, though Intel has not announced a broad commercial launch date. Tom’s Hardware reports that the company is collaborating with a wide ecosystem of OEMs, ODMs, cloud providers, and independent software vendors to integrate the hardware. This partnership approach aims to ensure software compatibility and broader industry adoption ahead of general availability.
Background: Large language models
Large language models are statistical engines that predict the next token based on patterns in their training data. They do not "know" facts; they generate plausible continuations. You must verify their outputs, secure their inputs against injection, and manage their context windows carefully to prevent data leakage.
Read the full guide: Large Language Models: 8 Technical FAQs for Secure Development
What to do and how to stay safe
- Monitor vendor announcements for detailed software stack compatibility with existing AI frameworks.
- Evaluate current data center cooling capacity to determine if air-cooled solutions fit infrastructure constraints.
- Review total cost of ownership models for inference workloads against liquid-cooled alternatives.
- Engage with hardware partners
Step-by-step guide: How AI Risk Management Works: The Hidden Failure Points
General security guidance from the Sandbox Dispatch newsroom. It is not confirmed advice from the organisations named in this story.
