# Hardware Feature required for DoCA to work

**URL:** <https://forums.developer.nvidia.com/t/hardware-feature-required-for-doca-to-work/355328>\
**Category:** Getting Started & Resources\
**Created:** [December 22, 2025, 8:04am UTC](https://forums.developer.nvidia.com/t/hardware-feature-required-for-doca-to-work/355328 "2025-12-22T08:04:25Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![palok](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@palok](https://forums.developer.nvidia.com/u/palok)\
**Post date:** [December 22, 2025, 8:04am UTC](https://forums.developer.nvidia.com/t/hardware-feature-required-for-doca-to-work/355328/1 "2025-12-22T08:04:25Z")

</div>

I am trying to understand which NVIDIA/Mellanox NIC will work seamlessly with DOCA GPUNetIO  
and what features in them maps .  
Also in paper

> **[GPU-Initiated Networking for NCCL](https://arxiv.org/html/2511.15076v1)**

”For GPU-initiated communication, the hardware requirement is NIC support for device-accessible control structures. NVIDIA ConnectX series adapters (ConnectX-6 Dx and later)”  
How NIC has enabled suport for device accessible control structure to GPU?

---

<div class="post-metadata">

**Author:** ![zhangsuo](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@zhangsuo](https://forums.developer.nvidia.com/u/zhangsuo)\
**Post date:** [January 16, 2026, 9:41am UTC](https://forums.developer.nvidia.com/t/hardware-feature-required-for-doca-to-work/355328/2 "2026-01-16T09:41:35Z")

</div>

> [@palok](#):
>
> DOCA GPUNetIO

Hi,

NVIDIA recommends using ConnectX-7 or later adapters. ConnectX6-Dx still works but some features (e.g., Ethernet Accurate Send Scheduling) are not supported.

- **[GPUDirect RDMA](https://docs.nvidia.com/cuda/gpudirect-rdma/index.html)** – Enables direct packet transfer between the NIC and GPU memory, eliminating unnecessary memory copies

- **[GPUDirect Async Kernel-Initiated](https://ieeexplore.ieee.org/document/7973709)** (GDAKI) – Allows CUDA kernels to control network operations without CPU intervention

- **[GDRCopy](https://github.com/NVIDIA/gdrcopy)** Library – Allows the CPU to access GPU memory directly

- NVIDIA BlueField DMA Engine – Supports GPU-triggered memory copies

For more, please refer:

> **[DOCA GPUNetIO - NVIDIA Docs](https://docs.nvidia.com/doca/sdk/doca-gpunetio/index.html)**

Thanks,

Suo