E810-XXVDA2 will not link at 25G with SFP28 optics until NVM is updated to 4.20
A brand-new retail Intel E810-XXVDA2 on Windows Server 2022 (Intel driver 1.12.164.0 from release package 28) was connected to an SFP28 port of a Ubiquiti USW-Pro-Aggregation switch. With an SFP28 DAC the link came up at 25G immediately, but with SFP28 optical transceivers (Ubiquiti UACC-OM-SFP-28-SR and a generic Intel-coded SFP28) no link was established at 25G; the same optics linked when forced to 10G, and SFP+ 10G optics worked as well. Port-to-port loopback at 25G passed on both the switch and the NIC, so the optics and fibre were fine. Forcing 25G on both ends and trying different FEC settings (the switch uses FC-FEC/BASE-R and can also disable FEC) changed nothing. The adapter's driver properties also left the Media Type and Negotiated Link Speed fields empty.
- Hardware: Intel Ethernet Network Adapter E810-XXVDA2, Ubiquiti USW-Pro-Aggregation
- Transceivers: Ubiquiti UACC-OM-SFP-28-SR, generic Intel-coded SFP28 SR, SFP28 DAC, SFP28 AOC, third-party SFP+ 10G optics
- Form factor: SFP28, DAC, AOC, SFP+
Symptoms / messages:
no link at 25G with SFP28 optical transceivers
same optics link at 10G; SFP28 DAC links at 25G
port-to-port loopback at 25G passes on both switch and NIC
Media Type and Negotiated Link Speed blank in Windows driver properties
forcing 25G and changing FEC has no effect
Comments 2
Accepted answer from the poster: updating the adapter NVM with Intel's NVMUpdate tool (1.39.32.6) from 3.20 to 4.20 (released 17 Feb 2023, DevId 159B) fixed it instantly. With speed and FEC left on auto the link came up at 25G with every SFP28 module tried: the Ubiquiti UACC-OM-SFP-28-SR, the Intel-coded generic module, an FC-FEC SFP28 and an SFP28 AOC. After the update Media Type showed "Fiber-SR" and Negotiated Link Speed "16.0 GT/s".
Intel support asked for the transceiver model and whether autonegotiation had been disabled with 25G full duplex forced on both sides, then escalated the case as a suspected switch/NIC compatibility issue; the poster later noted that support never spotted the outdated NVM during more than two months of troubleshooting.