Hi NVIDIA team,
First, thank you for granting the Rivermax development license and for the support so far — it’s much appreciated.
I’m doing academic research on object tracking and segmentation. As part of the pipeline I need to receive a high-bandwidth video stream (~8 Gb/s, over fiber or RJ45) with kernel bypass and move the payloads directly into GPU memory via GPUDirect (zero-copy), so I maximize GPU utilization during ingest and keep the CPU out of the data path. I have two target platforms:
- An x86_64 workstation with an RTX 5080 (Linux).
- A Jetson AGX Thor Developer Kit (T5000, JetPack 7.x, Ubuntu 24.04, kernel 6.8-tegra) for edge deployment.
From the Rivermax adapter selection guidelines I understand Rivermax only binds to supported ConnectX-6 Lx/Dx, ConnectX-7, or BlueField-3 adapters. Neither machine currently has one:
- On the 5080 workstation,
lspcishows only a Realtek 2.5GbE controller and a MediaTek Wi-Fi adapter. - On Thor, the QSFP28 (4x25GbE) interfaces use the
nvethernet(6.8.12-tegra) driver and the RJ45 is a Realtek RTL8126 — neither is a ConnectX ASIC.
My questions:
- RTX 5080 workstation: To use Rivermax here, do I simply need to install a supported ConnectX NIC (e.g. ConnectX-6 Lx, 2x25GbE) into a free PCIe slot? For ~8 Gb/s, is a single ConnectX-6 Lx (PCIe 4.0 x8) sufficient? I also read that on Linux GPUDirect works on all GPUs — can you confirm it works with the RTX 5080 on Linux?
- Jetson AGX Thor: Is there any supported way to run Rivermax on Thor? Since its integrated networking is not ConnectX, is it feasible to attach a ConnectX-6/7 add-in card via an M.2-to-PCIe adapter, and do the DOCA-OFED / mlx5 drivers build and load on the JetPack 7.x Tegra (aarch64) kernel? Is Rivermax validated in that configuration?
- If Rivermax is not viable on Thor, then what would be recommended as an alternative path for kernel-bypass-to-GPU?
Thanks again for the support.