Skip to navigation
User GuideNIXL Backends

DOCA GPUNetIO

View as Markdown

Overview

The DOCA GPUNetIO backend provides high-performance GPU-to-GPU transfers using GPUDirect Async over the DOCA networking stack on supported NVIDIA SmartNICs.

PropertyValue
Transfer TypeVRAM ↔ DRAM
ProtocolDOCA GPUDirect Async
Best ForUltra-low-latency GPU-to-GPU transfers on DOCA-capable systems

Installation

Prerequisites

The DOCA GPUNetIO backend requires:

  • DOCA SDK — Version 3.5.0 or later, from the NVIDIA DOCA SDK page. Older DOCA releases use an incompatible GPUNetIO device verbs API. When doca-gpunetio reports a lower version the plugin is skipped, or, if you enabled it explicitly with -Denable_plugins=GPUNETIO, configuration fails.
  • GPUDirect Async-capable hardware — NVIDIA BlueField SmartNICs (DPUs)
  • CUDA Toolkit — Version 12.8 or later. This is NIXL’s own floor for the plugin; DOCA 3.5 itself is built for CUDA 13.

For system configuration and setup details, see the DOCA GPUNetIO Programming Guide.

Configuration

The DOCA GPUNetIO backend has no backend-specific environment variables or build options.

When to Use

  • Ultra-low-latency GPU-to-GPU transfers
  • Deployments with DOCA-capable NVIDIA SmartNICs (BlueField DPUs)
  • Workloads preferring the GPUDirect Async path over UCX RDMA