Intel outlined its Intel GPU D-series at the Open Compute Project Global Summit, describing an AI inference accelerator with up to 480GB of memory. The company is targeting enterprise AI services that need to accommodate large models, lengthy prompts and simultaneous requests, with a design intended for existing data centers.
“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, AI Products & GTM, Intel Data Center Group.
“With Intel GPU D-series, we’re focused on making high-capacity inference practical to deploy and economical to scale.”
Memory capacity and model-size claims
Intel specifies LPDDR5X memory for the D-series, with a stated capacity of up to 480GB per accelerator. The company presents that memory configuration as part of its approach to serving demanding AI inference workloads.
Intel says a single card can support large language models containing up to 300 billion parameters. It also claims an eight-GPU system can accommodate models exceeding one trillion parameters. These model-size claims have not been independently verified.
The intended applications include enterprise AI services handling long prompts and multiple requests at the same time. Intel also identifies robotics applications requiring responsive processing as a target, extending the proposed uses beyond systems that generate responses to text prompts.
The memory capacity is relevant to workloads that need to keep substantial model weights and associated data accessible during inference. In that setting, the amount of memory available determines how much of that material the accelerator can hold.
Intel rates the D-series at 1.3 petaflops for FP8 computation. The company also lists memory bandwidth of 1.5TB/s. These are stated hardware specifications; they do not establish independently verified performance for the enterprise services and robotics applications Intel is targeting.
The model-size claims, compute rating and memory specification describe different aspects of Intel’s proposed accelerator. The company’s stated emphasis is on making high-capacity inference practical to deploy and economical to scale, as Nanduri described, with support for both single-card and larger configurations.
Deployment in existing data centers
Intel specifies a 350W power envelope and a standard PCIe form factor for the D-series, alongside air cooling. The company intends these characteristics to simplify deployment in existing data centers and reduce the need for specialized cooling arrangements or new rack infrastructure.
That deployment approach is part of Intel’s positioning for enterprise inference. The proposed combination brings its high-memory accelerator into a conventional server expansion format, with air cooling included in the design rather than presented as a separate deployment option.
Intel also describes software compatibility across its GPU platforms as part of the offering. Alongside the accelerator, the company plans to use Xeon processors to coordinate tasks, manage data movement and handle workloads that do not require GPU acceleration.
The CPU and GPU roles fit the multistage AI systems Intel is addressing. Those systems can include retrieval, reasoning, orchestration, external tools and execution, with different processing requirements across the stages. Intel’s proposed arrangement assigns coordinating work and other suitable tasks to CPUs while using the accelerator for GPU workloads.
“Open models are expanding the options enterprises have for deploying AI,” said Nicolas Dubé, corporate vice president, Data Center Systems and Solutions Group, Intel.
“To realize their full potential, organizations need infrastructure that can match different workloads with the right intelligence, right compute resources and right economics.”
