If your Proxmox cluster runs Ceph on NVMe drives, a 10GbE network is probably already the slowest part of it. Proxmox’s own guide says a single NVMe drive can saturate 10 Gbps and recommends at least 25 Gbps for that kind of setup, and its benchmark showed the same ceiling. The question is which 25GbE card to put in each node, and the answer is awkward: of the four cards we pick, three are end of life or under an end-of-life notice.
We checked each card against NVIDIA, Intel and Broadcom documents, the Proxmox VE 9.2 kernel and firmware packages, and the Ceph docs on October 20, 2026. We don’t own 25GbE hardware, so every figure here is a datasheet value or a named third-party measurement. If 10GbE is still enough for you, our 10GbE NIC guide for Proxmox covers that tier and the install steps.
| Card | PCIe | Power on DAC (typ / max) | Airflow needed | Linux driver | RDMA | Status | Amazon, Oct 20 |
|---|---|---|---|---|---|---|---|
| NVIDIA ConnectX-4 Lx MCX4121A-ACAT | Gen3 x8 | 9.6 / 11.6 W | 300 LFM | mlx5_core | RoCE | End of life | about $85 |
| Intel E810-XXVDA2 | Gen4 x8 | 8.9 / 9.8 W (8.0 W idle) | 100 LFM | ice | iWARP, RoCEv2 | EOL notice, last order Jan 2027 | about $255 |
| NVIDIA ConnectX-6 Lx MCX631102AN-ADAT | Gen4 x8 | 11.53 / 12.94 W at Gen4 | 300 LFM | mlx5_core | RoCE | Current | about $449 |
| Intel XXV710-DA2 | Gen3 x8 | 8.6 / 11.8 W | 100 LFM | i40e | None | EOL notices (controller, OEM SKU) | about $150 |
| Broadcom P225P | Gen3 x8 | 10.9 / 11.1 W | 50 LFM | bnxt_en | RoCEv2 | Active | No genuine listing |
Every card here has two SFP28 ports and falls back to 10GbE with SFP+ modules or DACs, so you can buy the card before the switch. Prices are reseller listings on Amazon US, so they move.
Where 25GbE pays off in a Proxmox Ceph cluster
Ceph sends every write across the network more than once, because each object is copied to the other replicas, so the cluster network carries more traffic than your VMs ask for. The Proxmox Ceph guide asks for at least 10 Gbps dedicated to Ceph, and at least 25 Gbps once you use NVMe. When you’re unsure, it suggests a layout of 25 Gbps or more for the cluster network, 10 Gbps or more for the public network and a separate 1 Gbps link for corosync.
Proxmox put numbers on it in its December 2023 Ceph benchmark. On three NVMe nodes, a two-link 10GbE mesh capped writes at about 1,680 MiB/s. The same three nodes on a two-link 25GbE mesh reached about 4,000 MiB/s with FRR routing and two OSDs per node, and about 2,200 MiB/s with RSTP, which blocks one of the mesh links. The Ceph hardware recommendations agree, saying clusters with substantial workload do well on 25 Gb/s and that NVMe nodes can saturate 10 or even 25 GE.
Skip 25GbE in these two cases, because the network will not be what slows you down:
- HDD OSDs: a WD Red Pro drive moves 164 to 287 MB/s, so 10GbE covers two to four drives per node, and 25GbE only shortens recovery.
- Small-block VM IOPS: latency limits them, not bandwidth, so a faster link changes little inside the VM.
If none of those apply and your nodes have NVMe OSDs, the rest of this guide is about which card to fit. Our Ceph homelab hardware guide covers the rest of the node.
Stock Ceph doesn’t use RDMA
Card vendors put RoCE and iWARP on the box, but stock Ceph does not use them. Its messenger defaults to plain TCP, and the Ceph network reference warns that other transports may be experimental with limited support. Red Hat Ceph Storage 8 says it does not currently support RDMA, and a Proxmox forum thread reports OSDs crashing within seconds on the RDMA transport, though a later reply there says it worked on Proxmox VE 8.0.
RDMA still matters if you also run NVMe over Fabrics, iSER, SMB Direct or DRBD over RDMA, so it can be a reason to prefer one card. For a Ceph-only cluster, treat it as a feature you won’t use, and choose on driver, power and price instead.
The PCIe slot decides which card fits
A dual-port 25GbE card needs 50 Gb/s, and the slot you have decides whether it gets that. A PCIe Gen3 x8 slot gives about 63 Gb/s, which is enough. A Gen3 x4 slot gives about 31.5 Gb/s, so one port runs at full speed but both together lose about 18.5 Gb/s before protocol overhead.
Gen4 cards solve that. The E810 and ConnectX-6 Lx get about 63 Gb/s from a Gen4 x4 slot, so both ports run at line rate in a mini PC or a board whose spare slot is only x4. A Gen3 card in that slot is fine if you use one port, for example one link to the switch. Narrow chipset slots cut it further. Gen4 x2 gives about 31.5 Gb/s, enough for one 25G port on a Gen4 card, while Gen3 x2 gives about 15.75 Gb/s and a Gen3 x1 slot drops below 10 Gb/s. Check the slot width on Minisforum and similar mini PCs before you buy.
ConnectX-4 Lx, the value pick
The MCX4121A-ACAT was the cheapest genuine retail listing for two 25GbE ports, at about $85 on October 20. It uses the in-tree mlx5_core driver, needs no extra firmware file on the host, and supports 127 SR-IOV virtual functions per port. NVIDIA has moved the ConnectX-4 Lx line to end of life, though its firmware release notes still list current firmware, so new stock is old stock.
Idle power is its weak spot. NVIDIA support says the card has no power-saving capabilities, and a Level1Techs measurement saw it add about 6 W at the wall on an N100 board with no link. On a W680 Proxmox host, a user on the NVIDIA forum saw it add 13 to 14 W even though the system stayed in C10 almost all the time. It also wants 300 LFM of airflow, per the user manual, which a desktop case does not give you.
Buy from a seller you trust, and check the card once it arrives. One reviewer on this listing received firmware NVIDIA called suspicious. Install mstflint with apt install mstflint, find the card with lspci -d 15b3:, then run mstflint -d 01:00.0 query with its address and look for the PSID MT_2420110034. A Dell, HP or Lenovo OEM board reports its own PSID, and one Amazon reviewer had to force-flash stock firmware onto a Dell card. The Vogzone board with the same chip was about $80, sold by Vogzone and shipped by Amazon, if you prefer an easy return.
Intel E810-XXVDA2, the best fit for Proxmox today
We would put the E810-XXVDA2 in a Proxmox cluster this year. It is PCIe Gen4 x8, so it keeps both ports at line rate even in a Gen4 x4 slot. Its product brief lists the lowest maximum DAC power here, 9.8 W, and the only published idle figure, 8.0 W. It needs only 100 LFM of airflow, supports iWARP and RoCEv2, and offers up to 256 SR-IOV virtual functions.
It is not a long-term part, though. Intel’s end-of-life notice from June 2026 sets the last order date for the retail card at January 1, 2027, with the E835-XXVDA2 named as its replacement. The driver stays in the kernel and the card keeps working, so buying one now is fine, but expect stock to thin out and prices to move during 2027.
The ice driver needs a DDP package, or the port drops to Safe Mode and, per the kernel docs, supports only basic traffic. On Proxmox you do not need to install anything, because the pve-firmware package ships the file. On plain Debian 13 it comes from the firmware-intel-misc package, and on Debian 12 from firmware-misc-nonfree.
Two forum issues matter before you build. A memory leak with MTU 9000 under heavy Ceph traffic was fixed in current Proxmox kernels, but a single report about LACP bonds on newer 6.17 builds and early 7.0 kernels was still open on October 20. The Intel retail card was about $255 from a reseller, and the 10Gtek E810 board was about $195, sold by 10Gtek and shipped by Amazon.
ConnectX-6 Lx, the current NVIDIA card
The ConnectX-6 Lx is the only one of our picks that is current with no end-of-life notice, so buy it if you want a part that stays in production. It moves to PCIe Gen4 x8, so it shares the slot advantage of the E810, and its user manual lists 11.53 W typical and 12.94 W maximum on DAC at Gen4. Like the ConnectX-4 Lx, it wants 300 LFM of airflow.
At about $449 on October 20, the genuine part costs five times as much as a ConnectX-4 Lx, which is hard to justify for a homelab. The LR-LINK board with the same chip was about $216, sold by LR-LINK and shipped by Amazon. We found no independent testing of these third-party boards, so keep the return window in mind.
Intel XXV710-DA2, the budget Intel card
The XXV710-DA2 is the cheapest Intel route, at about $150 on October 20, sold by a third party and shipped by Amazon. It runs PCIe Gen3 x8, draws 8.6 W typical on DAC and needs 100 LFM of airflow, and the i40e driver is in every current kernel. For a Ceph-only cluster, its lack of RDMA does not matter, because Ceph doesn’t use RDMA anyway.
Three things count against it. Intel has end-of-life notices out for select XXV710-AM2 controller SKUs, with a last order date of December 31, 2026, and for the OEM version of this adapter. Some 700-series adapters accept only Intel optics, per the i40e kernel docs, so buy Intel-coded modules. The i40e docs also say you cannot set the speed on XXV710-based devices, so it relies on autonegotiation with the switch.
The cards we left out
Three more cards came up in our research and did not make the picks:
- ConnectX-5 MCX512A-ACAT: end of life, and about $244 for the genuine part, more than a third-party ConnectX-6 Lx board.
- Broadcom P225P: active, low airflow need (50 LFM) and RoCEv2, but we found no genuine Amazon listing, only clones of its chip.
- Chelsio T6225-CR: iWARP and 13 W typical, but it accepts only Chelsio-supplied modules.
We also skipped several Amazon listings that looked wrong on inspection. One claimed a different controller from the one NVIDIA lists for the MCX4121A-ACAT, another was listed as x4 for a card NVIDIA rates at x8, and reviews on a third reported a malformed MAC address. Check the part number against the vendor document before you trust a listing.
Cables, optics and the switch side
Use DACs where you can, as our DAC vs fiber guide explains for 10G. They cost less than optics, and they draw less power: optics add 3 to 4 W on the E810, for example. These are the parts we would buy for a three-node cluster on October 20:
- SFP28 DAC, 1 m or 3 m: the 10Gtek 1 m and 3 m cables were about $16 and $20, sold by 10Gtek and shipped by Amazon.
- SFP28 SR optics: 10Gtek Intel-coded and Mellanox-coded modules were about $20 each. Match the coding to the card.
- QSFP28 to 4x SFP28 breakout: about $46 for the 10Gtek breakout DAC, which turns one 100G switch port into four 25G links.
A 25GbE card on a 10GbE switch runs at 10GbE, so the card buys nothing until the switch has SFP28 ports. The cheapest switch path is the MikroTik CRS504, with four QSFP28 ports that break out to 16 25G links. It lists at $799, and an Amazon listing from a third party was about $791 with a 9 to 10 day ship time. On Amazon US the cheapest with native SFP28 ports was the CRS510, at about $879 from a third party with one unit left. The UniFi USW-Pro-Aggregation lists from $899, but it was sold out at the Ubiquiti store. For exactly three nodes, you can skip the switch and run a full mesh of DACs, which is how the Proxmox benchmark wired its 25GbE cluster. Use the routed FRR setup, which the benchmark found clearly faster than RSTP.
The upgrade path from 10GbE
You don’t have to change everything at once. A sensible order keeps the cluster working at each step:
- Cards first: fit the 25GbE cards and run them at 10GbE on your existing switch and SFP+ cables.
- Ceph traffic next: move the Ceph cluster network to a three-node DAC mesh, or to one SFP28 switch. A mesh uses both ports of a dual-port card, so the public network then needs another NIC.
- The switch last: buy the SFP28 switch when you add a fourth node or want the public network at 25GbE too.
Each node also draws more power. Intel rates the E810 card itself at 8 W idle, and the ConnectX-4 Lx measurements above added 6 to 14 W at the wall depending on the board, so measure your own nodes before and after. For the rest of the build, see our Ceph cluster build guide, and keep corosync on its own 1 Gbps link, as Proxmox recommends, so a busy Ceph network never costs you quorum.
