--- 格式版本: 2 标题: "AMD CDNA™ Architecture" 原文链接: "https://www.amd.com/en/technologies/cdna.html" 发布日期: "2026-09-01" 发布时间校准状态: "found" 发布时间需复核: "否" 发布时间来源: "llm:scrape:original_meta_loose" 发布时间证据: "og:updated_time: 2026-09-01T02:24:12-0500" 发布时间校准原因: "候选日期来自页面元数据og:updated_time，属于页面发布时间/更新时间，符合发布时间判定优先级。" 发布时间校准置信度: "1" 发布时间候选数量: 2 发布时间严格候选数量: 0 发布时间原页读取状态: "source template page reused from URL open" 发布时间未找到原因: "" 发布时间校准时间: "2026-09-02T10:45:04+08:00" 发布时间仲裁状态: "confirmed" 发布时间仲裁尝试次数: 1 发布时间仲裁耗时毫秒: 7746 发现时间: "2026-09-02T10:31:49+08:00" 入库时间: "2026-09-02T02:45:15.515Z" 来源平台: "固定入口" 搜索渠道: "fixed_url" 搜索词: "https://www.amd.com/en/products/accelerators/instinct.html" 匹配关键词: - "Scale-up" - "HBM" - "performance" - "latency" - "bandwidth" - "throughput" - "GPU" - "AI" 相关厂家: - "AMD" 相关专家: [] 内容类型: "网页" 抓取工具: "Free Fetch + Defuddle" 清洗工具: "Defuddle Markdown + Defuddle/Readability 正文提取" 原始附件: [] AI优质: "是" AI打分: 89 AI分档: "高置信优质" AI质检状态: "通过" AI打分理由: "正文主线是AMD官方发布的CDNA 5、MI455X及72 GPU Helios机架级方案。相较知识库中7月的MI400/Helios发布，本文新增并集中披露2nm+3nm、最高3200亿晶体管、432GB HBM4与23.3TB/s带宽、12堆栈封装、Wave32/WGP、TDM与L2组播、每模块36组双向UALoE链路及PCIe 6/AI-NIC配置，并给出MI430X计划于2027年发布。当前页面是厂商一手架构资料，事实完整可追溯；虽缺少具名客户、订单和量产数据，但明确命中新架构/新产品及正式路线图通道。" AI质检模型: "gpt-5.6-sol" AI质检时间: "2026-09-08T16:44:07+08:00" AI主题相关性: 20 AI来源权威性: 15 AI新颖性: 17 AI技术细节: 20 AI商业部署信号: 7 AI完整性: 10 AI评分提示词版本: "v17-精简生产版" AI评分提示词SHA256: "48fb9777f386026761b4873eaff30807694fb11e9b352d7c69bf2dfde750cc7d" AI评分知识库版本: "knowledge_base_v1-20260819+runtime.92" AI评分知识库SHA256: "87cb146754f554920189197de0a266d3b3bbdd834b3853337f49da5d6a0d6c61" AI评分知识库检索词: "[\"AMD\",\"Scale-up\",\"HBM\",\"performance\",\"latency\",\"bandwidth\",\"throughput\",\"GPU\",\"HBM3E\",\"HBM4\",\"PCIe\",\"CDNA\"]" AI评分知识库命中: "[{\"id\":\"july-correct-0016\",\"title\":\"AMD launches Instinct MI400 Series GPUs for AI workloads\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"AMD\",\"Scale-up\",\"HBM\",\"performance\",\"latency\",\"bandwidth\",\"GPU\",\"HBM3E\",\"HBM4\",\"CDNA\"],\"rank\":-30.08929105479112},{\"id\":\"july-correct-0095\",\"title\":\"Inside NVIDIA Rubin GPU Architecture: Powering the Era of Agentic AI\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"Scale-up\",\"HBM\",\"performance\",\"latency\",\"bandwidth\",\"throughput\",\"GPU\",\"HBM3E\",\"HBM4\",\"PCIe\"],\"rank\":-23.678980110896358},{\"id\":\"runtime-4abbccc42d96af674efc7768\",\"title\":\"OpenAI’ Jalapeño: Better Than Nvidia Blackwell\",\"sourceType\":\"ai_excellent_article\",\"time\":\"2026-08-25\",\"matchedTerms\":[\"AMD\",\"Scale-up\",\"HBM\",\"performance\",\"latency\",\"bandwidth\",\"throughput\",\"GPU\",\"HBM3E\",\"HBM4\",\"PCIe\"],\"rank\":-21.51601557931416},{\"id\":\"july-correct-0079\",\"title\":\"AAI 2026: AMD Launches AMD Helios Rackscale Solution for Frontier AI\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"AMD\",\"Scale-up\",\"HBM\",\"performance\",\"latency\",\"bandwidth\",\"throughput\",\"GPU\",\"HBM4\"],\"rank\":-21.434059289263548},{\"id\":\"july-correct-0017\",\"title\":\"AMD challenges Nvidia’s networking dominance with Helios racks boasting 50% higher bandwidth\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"AMD\",\"Scale-up\",\"HBM\",\"performance\",\"latency\",\"bandwidth\",\"throughput\",\"GPU\",\"HBM4\"],\"rank\":-21.190728768657614}]" 采集批次: "2026年9月2日10点28分39秒" 采集批次ID: "20260902-102839-720" 去重键: "https://www.amd.com/en/technologies/cdna.html" --- ![](https://www.amd.com/content/dam/amd/en/images/backgrounds/products/5086250-cdna-top-row.jpg) ## Overview AMD CDNA™ architecture is the dedicated compute architecture underlying AMD Instinct™ GPUs and APUs. It features advanced packaging that unifies AMD chiplet technologies and High Bandwidth Memory (HBM), a high throughput Infinity Architecture fabric, and offers advanced Matrix Core Technology that supports a comprehensive set of AI and HPC data formats—designed to reduce data movement overhead and enhance power efficiency. Table Comparison between generations: | | **CDNA** | **CDNA 2** | **CDNA 3** | **CDNA 4** | **CDNA 5** | | --- | --- | --- | --- | --- | --- | | **Process Technology** | 7nm FinFET | 6nm FinFET | 5nm + 6nm FinFET | 3nm + 6nm FinFET | 2nm + 3nm FinFET | | **Transistors** | 25.6 Billion | Up to 58 Billion | Up to 146 Billion | Up to 185 Billion | Up to 320 Billion | | **CUs \| Matrix Cores** | 120 \| 440 | Up to 220 \| 880 | Up to 304 \| 1216 | 256 \| 1024 | 256 (WGP)\*\* | | **Memory Type** | 32GB HBM2 | Up to 128GB HBM2E | Up to 256GB HBM3 \| HBM3E | 288 GB HBM3E | 432 GB HBM4 | | **Memory Bandwidth (Peak)** | 1.2 TB/s | Up to 3.2 TB/s | Up to 6 TB/s | 8 TB/s | 23.3 TB/s | | **AMD Infinity Cache™** | N/A | N/A | 256 MB | 256MB | N/A | | **GPU Coherency** | N/A | Cache | Cache and HBM | Cache and HBM | Cache and HBM | | **Data Type Support** | INT4, INT8, BF16, FP16, FP32, FP64 | INT4, INT8, BF16, FP16, FP32, FP64 | **Matrix:** INT8, FP8, BF16, FP16, TF32, FP32, FP64 **Vector**: FP16, FP32, FP64 **Sparsity**: INT8, FP8, BF16, FP16 | **Matrix:** MXFP4, MXFP6, INT8, MXFP8, OCP FP8, BF16, FP16, TF32\*, FP32, FP64 **Vector:** FP16, FP32, FP64 **Sparsity:** OCP-FP8, INT8, FP16, BF16 | **OCP:** MXFP4, MXFP6, MXFP8, FP8 **Matrix:** INT8,BF16, FP16, FP32, FP64 **Vector:** FP16, FP32, FP64 **Sparsity:** INT8, FP16, BF16 | | **Products** | AMD Instinct™ MI100 Series | AMD Instinct™ MI200 Series | AMD Instinct™ MI300 Series | AMD Instinct™ MI350 Series | AMD Instinct™ MI400 Series | \*TF32 is supported by software emulation. \*\*AMD CDNA™ 5 architecture uses advanced Work Group Processors (WGP). ![](https://www.amd.com/content/dam/amd/en/images/backgrounds/products/5086250-cdna-5-sides.jpg) ![short top curved fade divider](https://www.amd.com/content/dam/amd/en/images/backgrounds/dividers/curved-gradient-divider-black-top-short.png) ## Introducing AMD CDNA™ 5 Architecture AMD CDNA™ 5 is the dedicated compute architecture underlying AMD Instinct™ MI400 Series GPUs. It features advanced packaging with chiplet technologies—designed to reduce data movement overhead and enhance power efficiency. The AMD Instinct™ MI455X GPU, designed specifically for the new 72 GPU AMD Helios™ rackscale solution, is the first product introduced based on the AMD CDNA 5 architecture with enhanced features for modern day AI scaling in frontier AI deployments. The AMD Instinct MI430X GPU for sovereign AI and HPC are planned for a 2027 release and may have different AMD CDNA 5 features enabled. ### AMD Instinct™ MI455X GPUs ![AMD Instinct™ MI455 Series GPU](https://www.amd.com/content/dam/amd/en/images/products/data-centers/4801352-mi455x-expanded-die-01.jpg) ### Advanced Packaging AMD CDNA 5 offers an advanced chiplet architecture that partitions compute, memory, cache, and I/O functions across specialized dies to optimize each function independently for performance and power. The 3D hybrid bonded compute dies are stacked and connected through high density die-to-die interconnects for improved compute density and efficiencies. AMD CDNA 5 integrates 432 GB of Next-Gen HBM4 memory in 12 stacks with 23.3 TB/s bandwidth and uses the AMD Infinity Fabric™ interconnect for high-bandwidth, low latency on-package communications between compute, memory, and I/O dies on the CoWoS-L advanced package to enable the memory and bandwidth scale required for modern AI scaling. ### Enhanced Compute AMD CDNA 5 offers a new Work Group Processor (WGP) architecture with higher throughput per clock to AMD Instinct™ MI400 Series GPUs. The new architecture enables Wave32 execution for reduced synchronization overhead and improved SIMD utilization and brings an advanced AI math engine with new tanh instructions and improved transcendental throughput. Advanced low-precision AI datatypes MXFP8, MXFP6, MXFP4 with block-scale 16/32 and fractional scaling of up to 4X the throughput compared to previous Gen AMD Instinct GPUs.1 ![AMD Instinct™ MI455 Series GPU](https://www.amd.com/content/dam/amd/en/images/products/data-centers/4801352-mi455x-expanded-die-03.jpg) ### Upgraded Memory System AMD Instinct MI400 Series GPUs raise the bar on memory capabilities offering an industry-leading 432GB HBM4 with 23.3 TB/s bandwidth per GPU and brings improved cache and memory hierarchy for growing models, context windows, and KV caches. AMD Instinct MI455X GPUs in the AMD Helios rackscale solution offer a 31TBs HBM4 shared pod memory system with 72 GPUs, all-to-all communications, single hop using an UALoE Fabric for large AI model workloads. ### Unified Fabric and I/O AMD Infinity Architecture, along with AMD Infinity Fabric™ technology, enable coherent, high-throughput on-package communication between AMD GPU chiplet technology, stacked HBM4 memory, L2 caches, and I/O dies in single devices and across multi-device platforms. Each AMD Instinct MI455X GPU EAM, designed for the 72 GPU AMD Helios rackscale solution, includes two enhanced I/O dies with either 2x PCIe® 6 compatible NICs or 3x AMD AI-NICs. Each GPU EAM also includes 36 (Bi-directional) UALoE links (x2) for high-bandwidth scale-up in AMD Helios rackscale solutions. ![AMD Instinct™ MI455 Series GPU](https://www.amd.com/content/dam/amd/en/images/products/data-centers/4801352-MI455X-expanded-die-06.jpg) ![Abstract background](https://www.amd.com/content/dam/amd/en/images/blogs/designs/generic-thumbnails/9851-adaptive-embedded-computing-04.jpg) ### Improved Efficiencies The AMD CDNA 5 architecture also brings improved GPU efficiencies by reducing memory traffic and idle cycles to increase delivered performance. Some key innovations include a Tensor Data Mover (TDM) for direct async global LDS that transfers without register staging, L2 multicast feature that allows one memory fetch to serve multi-WGPs to eliminate redundant memory traffic, split-named barriers for more efficient synchronization, and WGP clustering for multi-WGPs cooperation on Matrix operations. ![short bottom curved fade divider](https://www.amd.com/content/dam/amd/en/images/backgrounds/dividers/curved-gradient-divider-black-bottom-short.png) ## AMD CDNA™ 4 AMD CDNA™ 4 is the dedicated compute architecture underlying AMD Instinct™ MI350 Series GPUs. It features advanced packaging with chiplet technologies—designed to reduce data movement overhead and enhance power efficiency. ![AMD Instinct MI350 Series](https://www.amd.com/content/dam/amd/en/images/products/data-centers/3366850-instinct-mi350-closeup.png) ### AMD Instinct MI350 Series GPUs ![](https://www.amd.com/content/dam/amd/en/images/backgrounds/abstract/2325906-amd-cdna-grid-background.jpg) ![](https://www.amd.com/content/dam/amd/en/images/backgrounds/dividers/curved-gradient-divider-top.png) ## AMD CDNA 3 AMD CDNA 3 architecture is the dedicated compute architecture underlying AMD Instinct™ MI300 Series GPUs. It features advanced packaging with chiplet technologies—designed to reduce data movement overhead and enhance power efficiency. AMD Instinct MI300A APU AMD Instinct MI325X GPU ![](https://www.amd.com/content/dam/amd/en/images/backgrounds/dividers/curved-gradient-divider-bottom.png) ![Add Alt Text](https://www.amd.com/content/dam/amd/en/images/products/processors/2325906-amd-cdna-2-chip.png) ### AMD CDNA 2 AMD CDNA 2 architecture is designed to accelerate even the most taxing scientific computing workloads and machine learning applications. It underlies AMD Instinct MI200 Series GPUs. ### AMD CDNA AMD CDNA architecture is a dedicated architecture for GPU-based compute that was designed to usher in the era of Exascale-class computing. It underlies AMD Instinct MI100 Series GPUs. ![Add Alt Text](https://www.amd.com/content/dam/amd/en/images/products/processors/2325906-amd-cdna-1-chip.png) ![](https://www.amd.com/content/dam/amd/en/images/backgrounds/products/5086250-instinct-family-slab.jpg/jcr:content/renditions/5086250-instinct-family-slab-t.jpg) ## AMD Instinct Accelerators Discover how AMD Instinct GPUs are setting new standards for Generative AI, training, and HPC. ![](https://www.amd.com/content/dam/amd/en/images/backgrounds/abstract/2325906-amd-rocm-software-seamless-background.jpg/jcr:content/renditions/2325906-amd-rocm-software-seamless-background-m.jpg) ## AMD ROCm™ Software AMD CDNA architecture is supported by AMD ROCm™ software, an open software stack that includes a broad set of programming models, tools, compilers, libraries, and runtimes for AI and HPC solution development targeting AMD Instinct GPUs. Footnotes \*Some features described above may not be available for all AMD Instinct™ MI400 Series GPUs. 1\. Based on AMD Performance Labs calculations (June 2026) using an AMD Instinct™ MI455X GPU, peak theoretical precision performance (FP32, FP16, BF16, MXFP6, MXFP8, FP8, MXFP4 Matrix/Vector), compared to published specifications for AMD Instinct™ MI355X, MI350X, MI325X, MI300X, MI250X, and MI100 GPUs. Results may vary by system configuration and datatype. MI400-006