Intel 发布面向 AI 推理的 D 系列 GPU，首款 D270 提供 1.3 petaFLOPS FP8 算力
Intel Unveils D-Series GPUs for AI Inference < Semiconductor < 기사본문 - The Elec Inc.
Intel 在 OCP Global Summit 2026 上正式公布面向 AI 推理的 D 系列 GPU 详细规格，该系列此前代号 Crescent Island，是首个采用 Intel 最新 Xe3P GPU 架构的产品线。首款型号 D270 采用风冷设计和 PCIe 5.0 接口，无需专用机架或液冷即可装入数据中心，配备 32 个 Xe3P 核心和 256 个矩阵引擎，FP8 算力 1.3 petaFLOPS，功耗 350 瓦，支持 160GB 至 480GB LPDDR5X 内存，单卡可运行最高 3000 亿参数模型，八卡系统可支持最高 1 万亿参数模型，目前正向客户提供验证样品。Intel 同时推出面向智能体 AI 的机架设计，符合 OCP ORV3 标准，可容纳超过 46000 个 CPU 核心，基于 Xeon 6+ 的配置功耗为 133.45 kW。
正文
当前显示原文，中文全文译文尚未就绪。
JEONG IL JOO; Published 2026.10.14 18:19; Reply 0
Intel has officially unveiled its D-series graphics processing units (GPUs) for artificial intelligence (AI) inference.
Intel announced the detailed specifications of the D-series on Oct. 12 at the Open Compute Project (OCP) Global Summit 2026 in San Jose, California. Previously known by the codename Crescent Island, the lineup is the first to incorporate Intel's latest Xe3P GPU architecture.
The GPUs feature an air-cooled design and support the standard Peripheral Component Interconnect Express (PCIe) 5.0 interface. They can be installed directly in data centers without requiring dedicated racks or additional liquid cooling infrastructure.
The D-series was developed to improve cost efficiency in data centers running continuous AI inference workloads. It is designed to handle memory-intensive large language models (LLMs), long-context workloads, and simultaneous inference requests from multiple users.
The first model announced is the D270. Intel increased the number of compute cores by removing circuitry used for gaming graphics rendering and ray tracing. The GPU contains 32 Xe3P cores and 256 matrix engines, delivering 1.3 petaFLOPS of computing performance in 8-bit floating-point (FP8) operations with a power rating of 350 watts.
The GPU supports low-power double data rate 5X (LPDDR5X) memory capacities ranging from 160 gigabytes (GB) to 480GB. A single GPU card can run AI models with up to 300 billion parameters, while an eight-GPU system can support models with up to 1 trillion parameters. Intel is currently supplying validation samples to customers.
Intel also introduced a rack design optimized for agentic AI, which involves AI systems executing complex, multistep tasks and relies heavily on central processing unit (CPU) performance. The new rack complies with the OCP's Open Rack Standard Version 3 (ORV3) and can accommodate over 46,000 CPU cores. A rack configuration based on Intel's Xeon 6+ server CPUs has a power rating of 133.45 kilowatts (kW).
Intel's GPU platform utilizes common Xe software, allowing developers to reuse code built for existing GPU environments.
“AI inference requires infrastructure that delivers sufficient performance at an appropriate cost,” said Anil Nanduri, Intel’s Vice President and Head of Data Center AI Product Go-to-Market (GTM). “The D-series will help customers scale large-scale inference cost-effectively.”