Will vDPA support come to Connectx-7 400G?

Infrastructure & Networking Adapters and Cables Ethernet Adapter Cards
1

We’re looking to setup KVM with accelerated virtio through vDPA but this does not seem to be supported on the MCX75310AAS-NEA_Ax connect-x 7 400G NIC.

All steps taken as per documentation to add VF’s, unbind VF interfaces from mlx5_core driver, change NIC to switchdev mode and rebind VF’s back to mlx5_core driver show no output from:

vdpa mgmtdev show

We’ve recently been able to find an older Connectx-6 DX card (MCX623106AC-CDA_Ax) and upon taking the same steps as above are able to see the following from vdpa mgmtdev show:

pci/0000:40:00.2:
supported_classes net
max_supported_vqs 65
dev_features CSUM GUEST_CSUM MTU HOST_TSO4 HOST_TSO6 STATUS CTRL_VQ CTRL_VLAN MQ CTRL_MAC_ADDR VERSION_1 ACCESS_PLATFORM
pci/0000:40:00.3:
supported_classes net
max_supported_vqs 65
dev_features CSUM GUEST_CSUM MTU HOST_TSO4 HOST_TSO6 STATUS CTRL_VQ CTRL_VLAN MQ CTRL_MAC_ADDR VERSION_1 ACCESS_PLATFORM
pci/0000:40:10.1:
supported_classes net
max_supported_vqs 65
dev_features CSUM GUEST_CSUM MTU HOST_TSO4 HOST_TSO6 STATUS CTRL_VQ CTRL_VLAN MQ CTRL_MAC_ADDR VERSION_1 ACCESS_PLATFORM
pci/0000:40:10.2:
supported_classes net
max_supported_vqs 65
dev_features CSUM GUEST_CSUM MTU HOST_TSO4 HOST_TSO6 STATUS CTRL_VQ CTRL_VLAN MQ CTRL_MAC_ADDR VERSION_1 ACCESS_PLATFORM

DPDK documentation only lists the connect-x 7 200G variant as supported as of the latest 25.03 release but why not the 400G option?

Are there plans to add this functionality to Connectx-7 400G? What alternative to vdpa for accelerated virtio is available with this NIC?

2

Hi bvd,

Thanks for posting your inquiry to the NVIDIA Developer Forums.

It is unclear to me why vDPA support would be relegated to certain ConnectX-7 SKUs (perhaps the DPDK documentation is only referencing which SKUs were tested end-to-end at release time).

When using vDPA, the control path is handled by the vDPA driver (mlx5 vDPA driver if userspace, or kernel vDPA driver for kernel vDPA). Datapath (translation of virtio descriptors to mlx5 descriptors and vice versa) is offloaded to the NIC. Irrespective of individual SKU, I would expect the implementation would be the same (or similar) across this generation of adapter.

This will require a deeper dive than we’re able to satisfy in the community forums. If you have NVIDIA Enterprise support entitlement, we highly recommend opening a support case with your configuration and output so we can assist you further.

Thanks,
NVIDIA Enterprise Experience