We are building a commercial IaaS+PaaS cloud SDN data plane using DOCA Flow
on ConnectX-7. We need IPv6 Connection Tracking which requires the
format_select_dw_8_6_ext (full_dw_jumbo_support) capability bit.
We have tested three different SKUs:
- MCX75310AAS-NEA (PSID MT_0000000838, 400G NDR OSFP)
- MCX713106AS-VEA (PSID MT_0000000840, 200G QSFP112)
- MCX755206AS-NEA (PSID MT_0000000892, DGX storage)
All three: format_select_dw_8_6_ext = 0 in BOTH GET_CUR and GET_MAX
on firmware 28.48.1000 (latest).
Steps we have already verified:
- No mlxconfig variable affects this bit (tested 24 NV-config changes
with mlxfwreset, all confirmed via -e -a query showing Current column). - CR-space direct access returns 0xbadacce6 (blocked on standalone CX-7).
- SET_HCA_CAP from userspace is filtered by kernel mlx5 driver (EINVAL).
- All public CX-7 PSIDs share identical MAIN_CODE section.
Questions:
- Is there a roadmap to enable IPv6 CT pipe on standalone ConnectX-7?
- Is there a documented mechanism (signed CS/CRCS/DBG token) to unlock
this capability for evaluation/research? - The DOCA Flow CT documentation states “BlueField-3 and above required
for IPv6”. Is this a permanent product positioning decision or
firmware/HW limitation?
We are happy to share our CRCS challenge tokens with NVIDIA Engineering
support team. We have an active deployment of 3+ CX-7-based SDN nodes
and would prefer not to migrate to BlueField-3 if there is a path forward.