Transceiver compatibility on NICs: driver whitelists

On a switch, module policy is a line in the network OS; on a server it is a decision made by the NIC driver and firmware, and it differs per vendor, per operating system and per driver version. Intel adapters are the ones that refuse unlisted modules; most others accept anything that parses. This page lists the behaviour by vendor and OS, the documented overrides, the FEC/auto-negotiation rules that decide whether DAC and optics link, and what a coded identity must satisfy on the server side.

Behaviour by vendor

Vendor / driver	Check	On failure	Override	Notes
Intel ixgbe (X520, X540, X550 SFP+)	vendor OUI / PN list of Intel-qualified SFP+	`failed to load because an unsupported SFP+ or QSFP module type was detected`, port down	module parameter `allow_unsupported_sfp=1` (Linux); reload driver	DAC generally accepted; 1G SFPs limited
Intel i40e (X710, XXV710, XL710)	firmware-level qualification of the module	`unsupported module` in dmesg, link down or "module not qualified" in `ethtool`	`allow_unsupported_sfp=1` on newer i40e; older builds none	XXV710 25G modules strictly checked
Intel ice (E810)	NVM/firmware qualification	`Possible mis-configuration of the Ethernet port detected` / unqualified module	depends on NVM and driver release; some builds honour an unsupported-SFP setting, others reject — check Intel's release notes	most permissive of the Intel line in recent NVM
Intel on Windows (PROSet drivers)	same lists	device shows "Network cable unplugged", event log warning	no documented override	use Intel-listed or Intel-coded modules
Intel on VMware ESXi (ixgben, i40en, icen)	same lists	link down, `vmkernel.log` "unsupported module"	module parameters via `esxcli system module parameters set` where the driver exposes `allow_unsupported_sfp`; not on all versions	prefer qualified modules in production
Nvidia / Mellanox mlx5 (ConnectX)	parses EEPROM; Ethernet: no list	link up; `mlxlink` shows module info	not needed	InfiniBand mode: fabric manager validates cable/module identity and may flag or downgrade (IB cables)
Broadcom bnxt_en (NetXtreme-E)	parses; no list on most SKUs	up	—	OEM-branded (Dell, HPE) firmware may add checks
Marvell / QLogic qede	parses	up	—	—
AMD Solarflare sfc	parses	up	—	—
Chelsio cxgb4	parses	up	—	—
Pensando ionic	parses	up	—	—
Server-vendor OEM NICs (Dell, HPE, Lenovo, Cisco UCS VIC)	may add their own lists, esp. Cisco VIC	"unsupported transceiver" in CIMC/UCSM	none officially	UCS VIC accepts Cisco-coded modules

Behaviour changes with driver and firmware (NVM) versions; verify on the exact combination (Compatibility matrices & firmware).

Applying the Intel override (Linux)

# temporary (until reload)
modprobe -r ixgbe && modprobe ixgbe allow_unsupported_sfp=1
# persistent
echo "options ixgbe allow_unsupported_sfp=1" > /etc/modprobe.d/ixgbe.conf
echo "options i40e allow_unsupported_sfp=1"  > /etc/modprobe.d/i40e.conf
dracut -f   # or update-initramfs -u
# verify
cat /sys/module/ixgbe/parameters/allow_unsupported_sfp
dmesg | grep -iE 'sfp|module|unsupported'
ethtool -m eth0

For multi-port cards the parameter takes a comma-separated list (one value per port) on some drivers. The setting is not supported by Intel for production; it removes the check, not the electrical requirements.

FEC and auto-negotiation on NICs

Link type	AN	FEC	Notes
25G DAC	on (Clause 73) both ends	negotiated (RS or BASE-R per cable class)	most "no link on DAC" cases are AN off on one end
25G SR / LR optics	off; forced 25000	RS-FEC for SR (required), LR: RS or none — must match switch	`ethtool --set-fec eth0 encoding rs`
100G DAC	on	RS-FEC	—
100G SR4 / CWDM4 / DR	off	RS-FEC required	—
100G LR4	off	none (RS on some platforms)	match the switch (FEC & AN)
200G / 400G	off for optics	RS(544) KP4 always	—
10G	off	none	—

Commands per OS: NIC tools and diagnostics; switch side: Port configuration recipes.

What a coded module must satisfy for a NIC

Target	Minimum
Intel ixgbe/i40e without override	vendor name and PN from Intel's qualified list (Intel-branded `E10GSFPSR`/`E10GSFPLR`/`E25GSFP28SR` or the OEM parts they resell), consistent compliance codes and checksums
Intel with override, all permissive vendors	valid checksums and compliance codes — no OEM coding needed
Cisco UCS VIC	Cisco-style identity as for Cisco switches (How each NOS validates a module)
InfiniBand HCAs	correct cable/module identity per IBTA and the fabric vendor's list

Coding guidance: Vendor lock & coding, Part numbering.

Typical NIC-side symptoms

Symptom	Likely cause	Check
Port absent or "unsupported module" in dmesg	driver whitelist	`dmesg`, `ethtool -m`, override or coded module
Link down, module recognised, good levels	FEC/AN mismatch, speed forced wrong	`ethtool eth0`, `ethtool --show-fec eth0` (Speed & rate)
Link up at wrong speed	auto picked 10G on a dual-rate module	force speed
Link up, low throughput	PCIe width/speed	`lspci -vv` (Ports & PCIe)
Module hot, temperature alarms	card in exhaust airflow	DDM via `ethtool -m`; move card or improve airflow
Works in Linux, not in Windows/ESXi	different driver policy	use qualified module for that OS
InfiniBand link at reduced width/speed	cable not on the list or too long	`ibstat`, `mlxlink`

In CodingBox

Server-side rejections are diagnosed from the same bytes as switch-side ones. Read vendor name, PN, OUI, compliance codes and checksums on Check transceiver; for Intel-only fleets the code database can hold the Intel identities that the drivers accept.