What is the NVIDIA DOCA Software Platform?

The NVIDIA DOCA platform is the unified software foundation designed to unlock the full potential of NVIDIA BlueField DPUs and NVIDIA ConnectX SuperNICs for agentic AI infrastructure. It provides microservices, libraries, drivers, and orchestration frameworks to build and operate accelerated agentic AI factories.

What is a DPU, and how does DOCA leverage it?

A data processing unit (DPU) is a new class of programmable processor designed to offload, accelerate, and isolate data-centric tasks — such as networking, security, and storage — from a server's main CPU. NVIDIA DOCA is the software framework that unlocks the full capabilities of these DPUs, functioning similarly to how CUDA allows programming of GPUs.

A SuperNIC is a specialized class of network accelerator designed specifically to handle the high-speed, demanding networking needs of artificial intelligence (AI) workloads in agentic AI factories. Unlike traditional NICs designed for general-purpose computing, SuperNICs are purpose-built to accelerate data transfers between thousands of GPUs, preventing bottlenecks during large-scale AI training and inference.

What is the difference between DOCA microservices, libraries, and drivers?

DOCA is a complete software platform with three distinct layers. DOCA Microservices are production-ready, containerized infrastructure services — such as DOCA Argus and DOCA SNAP — that can be pulled from NVIDIA NGC and deployed directly without custom development. DOCA Libraries are programmable APIs that give developers direct access to BlueField hardware capabilities — such as packet processing, congestion control, and GPU networking — for building custom, high-performance infrastructure services. DOCA Drivers provide the foundational networking, security, and storage interfaces that connect the platform to BlueField and ConnectX hardware.

Which NVIDIA hardware platforms are compatible with DOCA?

NVIDIA DOCA is the unified software platform for both NVIDIA BlueField DPUs and NVIDIA ConnectX SuperNICs.

How does DOCA accelerate multi-tenant networking for agentic AI?

DOCA accelerates multi-tenant networking for agentic AI through DOCA HBN, DOCA Weave, and DOCA Flow, offloading virtual switching, routing, tenant isolation, and packet processing from the host CPU to NVIDIA BlueField DPUs and ConnectX SuperNICs — delivering line-rate networking at up to 800 Gb/s while ensuring each AI workload operates in a secure, isolated network environment.

What is RDMA and how does DOCA accelerate it?

RDMA (Remote Direct Memory Access) enables servers to exchange data directly between memory over a network without involving the CPU — reducing latency and freeing compute resources for AI workloads. DOCA accelerates RDMA through DOCA GPUNetIO, which allows CUDA kernels to initiate RDMA operations directly from the GPU, bypassing the CPU and operating system networking stack entirely. This enables true overlap of network communication and GPU computation, delivering the low-latency, high-throughput data transfers that distributed AI training and inference depend on.

What role does DOCA play in AI-native storage?

DOCA enables AI-native storage through DOCA SNAP, which allows BlueField DPUs to emulate local NVMe storage devices for the host, and DOCA Memos, which accelerates KV cache movement and sharing across GPU memory, host memory, and storage tiers at pod scale. DOCA Vault extends this with in-silicon, zero-trust file access enforcement, ensuring only authorized AI workloads and agents can access sensitive data, independent of the host OS or storage system.

What is DOCA Memos and how does it benefit long-context AI inference?

DOCA Memos is a library for efficient KV cache movement, storage, and sharing at pod scale. It eliminates redundant recomputation by managing context memory across GPU memory, host memory, and storage tiers — reducing inference latency and improving GPU utilization across multi-node large language model deployments.

How does DOCA enhance cybersecurity with in-silicon protection?

DOCA enhances cybersecurity through hardware-enforced protection independent of the host OS. DOCA Argus delivers agentless threat detection up to 1,000x faster than existing solutions, DOCA Vault enforces zero-trust file access in silicon, and DOCA Flow enables line-rate network policy enforcement at up to 800 Gb/s — shifting security from reactive host-based defenses to continuous inline enforcement.

What is DOCA Flow used for in network security?

DOCA Flow enables hardware-accelerated packet processing and line-rate network policy enforcement at up to 800 Gb/s directly on BlueField DPUs and ConnectX SuperNICs — without CPU involvement. In network security, it is used to enforce traffic segmentation, isolate tenants and AI agents, inspect packets inline, and block unauthorized network activity before it reaches the host. This makes it the enforcement layer for network security policies across distributed AI factory environments.

What is DPF and how does it simplify DPU management at scale?

DPF (DOCA Platform Framework) is a Kubernetes-native orchestration framework that provisions, manages, and scales BlueField DPUs as Kubernetes nodes. It automates firmware updates, lifecycle management, and DOCA service deployment across DPU fleets — enabling platform teams to manage thousands of DPUs using the same tooling they use for the rest of their infrastructure, without direct hardware programming.

How does DOCA Telemetry provide full-stack visibility?

DOCA Telemetry provides real-time visibility into BlueField DPUs and connected infrastructure by collecting and streaming metrics, events, and logs directly from the DPU data path through a single unified pipeline — integrating natively with standard observability platforms without custom instrumentation.

What are the benefits of DOCA SNAP for developers building storage appliances?

DOCA SNAP allows developers to emulate local Non-Volatile Memory Express (NVMe) and VirtIO storage devices on BlueField DPUs, enabling remote storage to appear local to the host without modifying drivers or applications — accelerating storage performance while offloading I/O processing from the host CPU.

What is the meaning of BGP EVPN in the context of DOCA networking?

BGP EVPN (Border Gateway Protocol Ethernet Virtual Private Network) is a standards-based control plane protocol. In the context of DOCA networking, it enables software-defined network segmentation, workload mobility, and multi-tenant isolation — allowing BlueField DPUs to enforce consistent network policy across dynamic, large-scale AI factory infrastructure without manual configuration.

How can I start developing custom applications with DOCA?

Developers can get started with DOCA at developer.nvidia.com/networking/doca, where you can find the DOCA SDK, documentation, and other resources.

How can I deploy pre-built DOCA microservices quickly?

Pre-built DOCA microservices are available on NVIDIA NGC and can be pulled and deployed as containers directly into your environment.

What is the path for deploying DOCA in a production environment?

DOCA supports two deployment paths. For pre-built services, DOCA microservices are available as containers on NVIDIA NGC and can be deployed directly into production environments using standard container orchestration tools. For custom development, the DOCA SDK provides the libraries, drivers, and APIs needed to build and deploy custom services on BlueField and ConnectX. DOCA Platform Framework (DPF) provides the Kubernetes-native orchestration layer for provisioning, managing, and scaling DPU fleets across production infrastructure.

Where can I find video sessions and blogs about the latest DOCA features?

You can explore the latest DOCA Developer Day sessions across networking, storage, and cybersecurity on NVIDIA On-Demand, and view all developer-focused blogs at developer.nvidia.com.

How does DOCA ensure my infrastructure platform is future-proof?

DOCA provides multi-generational support across both BlueField DPUs and ConnectX SuperNICs, ensuring applications built today run with added performance benefits on future hardware generations — without requiring code changes.

NVIDIA DOCA 软件平台

在 NVIDIA BlueField 和 ConnectX 上加速、保护和扩展代理式 AI 基础设施。

开始开发

概览
示意图
优势
工作负载
入门选项
构建
资源
常见问题解答
后续步骤

概览
示意图
优势
工作负载
入门选项
构建
资源
常见问题解答
后续步骤

开始使用

概览

什么是 NVIDIA DOCA？

NVIDIA DOCA™ 是一个统一的软件平台，可释放 NVIDIA® BlueField® DPU 和 NVIDIA® ConnectX® SuperNIC™ 的全部潜力，并赋能代理式 AI 基础设施。它提供 SDK、库、生产就绪型微服务和编排框架，用于大规模构建和运行加速的基础设施服务，包括网络、AI 原生存储、运行时安全、遥测和生命周期管理。

NVIDIA 推出由 DOCA 赋能的、面向代理式 AI 的内置安全存储

由 DOCA 赋能的 NVIDIA Vera BlueField‑4 STX 定义了一种面向代理式 AI 的新型内置安全存储，可用于数据、智能体和内存的保护。

阅读新闻稿

BlueField‑4 STX：由 DOCA 赋能的、面向代理式 AI 的安全存储

DOCA 为 BlueField-4 STX 提供在线智能体、数据和内存保护，确保代理式 AI 以基础设施速度安全运行。

阅读博客

示意图

NVIDIA DOCA 软件平台

NVIDIA DOCA 为基础设施团队提供所需的微服务、库和管理工具，以加速网络、强化运行时安全、管理 AI 原生存储，并保护基于 NVIDIA BlueField 和 NVIDIA ConnectX 的代理式 AI 工作负载。

查看 NVIDIA DOCA 堆栈

优势

专为满足代理式 AI 基础设施的需求而构建

NVIDIA DOCA 为开发者和构建者提供软件基础模块，助力他们基于 NVIDIA BlueField 和 NVIDIA ConnectX 创建加速且差异化的基础设施服务。它简化了开发流程、加速了部署速度、扩展了创新范畴，并为主机、网络和数据路径提供了深度可编程、可视性和生命周期管理。

加速开发

借助企业级 API、参考应用和即用型微服务，实现网络、安全、存储、遥测和基础设施编排，从而加快产品上市和创新。

随处部署

为 BlueField 和 ConnectX 创建应用程序，并借助原生 Kubernetes 集成、一致的操作和按需扩展，轻松部署到企业、云和边缘环境。

提升性能

通过 BlueField 和 ConnectX 上专为网络、安全和数据处理而构建的引擎，实现硬件加速、软件定义的控制。

面向未来的基础设施

只需一次构建，即可在 BlueField 和 ConnectX 各代产品之间无缝迁移，同时享受新硬件性能带来的优势。

工作负载

赋能代理式 AI 工厂操作系统的软件平台

NVIDIA DOCA 为运行 AI 工厂提供软件基础，使开发人员能够构建智能、安全且高性能的基础设施服务，并在复杂的分布式环境中无限扩展。

网络

加速代理式 AI 基础设施的多租户网络。

DOCA 微服务和库为代理式 AI 工厂提供高性能网络，在 BlueField 和 ConnectX 上加速虚拟交换、数据包处理、多租户隔离和高级流量管理。凭借超低延迟连接和高达 800Gb/s 的速度，DOCA 可实现大规模高效流量，同时 RDMA 加速助力最大化 GPU 利用率和集群效率。

DOCA 专为要求严苛的网络环境而构建，支持软件定义网络，包括分布式路由、Overlay 网络和 BGP EVPN，从而简化动态云环境中的网络分段、移动性和运维。通过将关键的网络处理从主机卸载，DOCA 提供了线速性能，同时释放了主机计算资源以承载 AI 服务和应用程序。

存储

为代理式 AI 构建上下文记忆存储和 AI 原生数据平台。

DOCA 支持为现代 AI 工厂创建 AI 原生的 KV 缓存和存储基础设施。借助 NVIDIA DOCA Memos™，开发者可以在 AI Pod 中构建代理式 KV 缓存共享框架，从而扩展有效的 GPU 内存容量，减少重复计算，并提升长上下文、多轮代理式推理的每秒 Token 数。

DOCA 还加速了新一代存储服务，包括基于 NVMe-oF、存储虚拟化和 BlueField 上的数据服务。借助 DOCA SNAP，开发者可以构建软件定义的存储设备和虚拟化 NVMe 解决方案，将存储处理从主机 CPU 卸载，同时随着代理式工作负载的扩展提供高性能的数据访问。

网络安全

以 AI 智能体的速度提供芯片级安全。

DOCA 为开发者提供软件构建模块，以便在 BlueField 上为现代 AI 工厂创建智能、硬件加速的网络安全服务。借助 DOCA Argus 和 DOCA Vault，开发者可以检测异常 AI 智能体行为、监控系统完整性，并通过隔离的硬件保护措施来保护凭证、密钥和敏感资产。

DOCA 还通过 DOCA Flow 实现高性能网络安全，直接在数据路径中进行流量检查、分段、策略执行、威胁缓解和加速加密。通过将安全功能与主机 CPU 隔离，DOCA 助力在云规模下提升保护能力，并降低开销。

编排

简化大规模 BlueField DPU 的管理。

DOCA 平台框架 (DPF) 简化了 BlueField 的大规模部署，并通过 Kubernetes 原生操作模式管理基础设施服务的整个生命周期。开发者和运维人员可以将网络、存储和安全服务作为云原生应用程序进行部署，并使用标准的 Kubernetes API 实现升级、策略控制、扩展和日常运维的自动化。

DPF 还支持可信主机和零信任运维模式，以加强租户、主机和基础设施服务之间的隔离。DPF 与 NVIDIA DSX 操作系统和领先的企业级 Kubernetes 发行版集成，可在私有云、服务提供商和大规模 AI 工厂环境中实现一致的运维。

可观测性

获得对 AI 基础设施的全栈可视性。

DOCA 提供对 BlueField 和 ConnectX 网络的实时可观测性，使开发者和运维团队能够深入了解 AI 基础设施的运行状况、性能和利用率。流式指标、事件、计数器和诊断信息有助于监控分布式 AI 工厂环境中的网络吞吐量、延迟、存储活动、安全服务和系统资源消耗情况。

借助精细的遥测数据和集群范围的洞察，团队可以及早发现异常、更快地排除瓶颈，并在事件影响工作负载或 GPU 利用率之前对其做出响应。通过与现有的监控和分析平台集成，DOCA 遥测有助于大规模优化基础设施效率、提高正常运行时间，并保持可预测的性能。

入门选项

NVIDIA DOCA 入门指南

无论您是在评估基础设施平台、试点新用例，还是扩展到全面生产环境部署，NVIDIA DOCA 都能为您提供一条清晰的路径，助您加速构建安全的 AI 基础设施。

下载并构建

下载 NVIDIA DOCA SDK 并访问 API、库和示例应用程序，即可开始在 BlueField 和 ConnectX 上进行构建。

开始体验

集成

从 NVIDIA NGC™ 拉取预构建的 DOCA 容器，即可在您的环境中快速部署可配置的基础设施微服务。

部署容器

查阅文档

浏览完整的 DOCA 文档库，包括 SDK 指南、API 参考以及 BlueField 和 ConnectX 的发行说明。

查看文档

在生产环境中部署

使用 NVIDIA AI Enterprise 支持的 DOCA，即可大规模部署可用于生产环境的企业级 BlueField。

探索 AI Enterprise

DOCA 开发者日回顾

了解 DOCA 在网络、存储和网络安全领域的最新会议内容。

观看 DOCA 开发者日会议

构建

构建全球 AI 基础设施

从超大规模云提供商到数据中心基础设施和网络安全创新者，全球领先的企业都在基于 NVIDIA DOCA 构建高性能软件平台，以满足客户对性能、安全性和扩展性的需求。

资源

关于 NVIDIA DOCA 的常见问题

NVIDIA DOCA 平台是一个统一的软件基础，能够释放 NVIDIA BlueField DPU 和 NVIDIA ConnectX SuperNIC 的全部潜力，并赋能代理式 AI 基础设施。它提供微服务、库、驱动程序和编排框架，用于构建和运行加速的代理式 AI 工厂。

数据处理单元 (DPU) 是一种可编程处理器，将网络、安全和存储等以数据为中心的任务从服务器的主 CPU 中卸载、加速和隔离。NVIDIA DOCA 是一个软件框架，它能够释放 DPU 的全部潜能，类似 CUDA® 实现 GPU 编程的方式。

SuperNIC 是一类专门设计的网络加速器，满足代理式 AI 工厂中 AI 工作负载的高速、严苛的网络需求。与专为通用计算设计的传统网卡不同，SuperNIC 专为加速数千个 GPU 之间的数据传输而构建，从而避免大规模 AI 训练和推理过程中出现瓶颈。

DOCA 是一个完整的软件平台，包含三个不同的层。DOCA 微服务是可直接部署的生产就绪型容器化基础设施服务，例如 DOCA Argus 和 DOCA SNAP，它们可以从 NVIDIA NGC 拉取并直接部署，无需定制开发。DOCA 库是可编程 API，使开发者能够直接访问 BlueField 硬件功能 (例如数据包处理、拥塞控制和 GPU 网络)，从而构建定制的高性能基础设施服务。DOCA 驱动程序提供连接平台与 BlueField 和 ConnectX 硬件的基础网络、安全和存储接口。

DOCA 平台框架 (DPF) 是一个 Kubernetes 原生编排框架，可将 BlueField DPU 作为 Kubernetes 节点进行配置、管理和扩展。它可自动执行固件更新、生命周期管理以及跨 DPU 集群的 DOCA 服务部署，使平台团队能够使用与管理其他基础设施相同的工具来管理数千个 DPU，而无需直接进行硬件编程。

DOCA 支持两种部署路径。对于预构建的服务，DOCA 微服务以容器形式在 NVIDIA NGC 上提供，并可以使用标准容器编排工具直接部署到生产环境中。对于定制开发，DOCA SDK 提供了在 BlueField 和 ConnectX 上构建和部署定制服务所需的库、驱动程序和 API。DPF 提供了 Kubernetes 原生编排层，用于在生产基础设施中配置、管理和扩展 DPU 集群。

DOCA 为 BlueField DPU 和 ConnectX SuperNIC 提供跨代支持，确保当前构建的应用程序在未来新一代的硬件中也能获得更佳的性能，而无需修改代码。

后续步骤

准备好利用 DOCA 改造您的基础设施了吗？

无论您是为企业评估 NVIDIA DOCA、在 BlueField 上部署微服务，还是构建定制应用程序，NVIDIA 都能为您提供入门指南。

探索 DOCA

与 NVIDIA DOCA 专家交流

联系 NVIDIA 网络专家，讨论您的 AI 基础设施用例，并获取将 DOCA 集成到您的技术栈中的指导。

联系我们

随时了解 DOCA 社区的最新动态

在您的订阅中获取最新的 NVIDIA DOCA 版本说明、开发者博客和活动公告。

随时了解最新动态

NVIDIA DOCA 软件平台