Quick answer

Swap the switch fan or buy a fanless one; use passive DAC under five meters and fiber past that. But before any of it, check your storage: a single hard drive gives you a quarter of 10GbE, and one SMB transfer rides one CPU core. Skip ConnectX-3.

On this page

By LK Wood IV · Published 2026-09-07 · ~8 min read · St. Louis County, MO

Ten gigabits per second is 1.25 GB/s on the wire. A single Seagate Exos X24 is rated at 285 MB/s at its outer tracks, and that is Seagate’s own best case rather than what you will see.

That gap is the article. The fans, the cables and the cards are all real problems, and I will get to each of them, but none of them is the reason your transfers are slow.

The fan is the first thing you notice and the first thing you fix

Fan noise is what you will register in the first hour of owning a 10GbE switch, and it is what you will deal with first. There is an aftermarket in quiet replacement fans for these switches, and it exists because everyone who buys one hits the same wall in the same week.

The swap works. I do not treat it as a compromise or as a mod you put up with — a fan specified for a datacenter row is the wrong part for a shelf in a room where someone sits, and replacing it is the correct configuration for where the switch actually lives.

If you have not bought yet, there is a shorter path: buy a switch that never had a fan. MikroTik lists cooling type Passive for both the CRS305-1G-4S+IN, with four SFP+ ports, and the CRS309-1G-8S+IN, with eight. Four 10G ports covers a NAS, a workstation and a hypervisor with one spare, which is most houses.

DAC under five meters, fiber past that

Passive DAC for short runs. It is cheaper, it runs cooler, it draws less, and there are fewer parts in the path that can fail.

Cisco’s 10GBASE SFP+ datasheet puts a number on “short.” Its passive Twinax cables are listed at 1, 1.5, 2, 2.5, 3, 4 and 5 meters; active Twinax starts at 7. Five meters is where the industry draws the line.

The power figures in the same document: 0.1W for a passive Twinax cable, 1W for an SFP-10G-SR module — and an optical link needs two of those modules, one at each end. Ten times the draw per port, plus two components that can fail independently of the cable between them. Inside a rack or across a room, take the DAC. Through a wall or across a house, take fiber and accept the transceivers.

The bottleneck is almost never the network

This is the part nobody says, and it is the reason most 10GbE upgrades disappoint.

Run the arithmetic against your own storage before you spend anything:

Device or interfaceCeilingShare of a 10GbE link
10GbE link1,250 MB/s100%
One Exos X24 (285 MB/s rated)285 MB/s23%
SATA III interface (6 Gb/s, 8b/10b)600 MB/s48%
Four Exos striping, perfect sequential1,140 MB/s91%

A single hard drive gives you under a quarter of the link. A SATA SSD cannot reach half of it, and that is the interface ceiling before the flash behind it has any say. Four drives reading perfectly sequentially finally approach the wire, and nothing reads perfectly sequentially — the six-bay rebuild math covers what happens to those numbers once parity and scattered blocks are in play.

Saturating 10GbE takes NVMe or a genuinely fast array. Most people adding 10GbE do not have one.

And then there is SMB

Even with the storage to feed it, a single file copy will sit well under the link, and the protocol is why. Microsoft’s own account of the problem: with one SMB session and no Multichannel, SMB opens a single TCP connection, and that connection gets affinitized to a single CPU core. One copy, one connection, one core. A faster switch does not sell you a second core.

SMB Multichannel is the fix. On an RSS-capable NIC it opens several TCP connections for the same session and spreads them across cores. It is on by default on Windows. On TrueNAS SCALE it is a toggle under Advanced Settings in the SMB service, and flipping it costs nothing.

Check both of those before you buy hardware. “Just add 10GbE” is the most common wrong assumption in home networking, and the two things it usually fails to fix are sitting in your NAS and in your file-sharing config.

ConnectX-3 is done

The used ConnectX-3 was the budget answer for years and it is now the wrong buy at any price.

VMware’s list of devices dropped in ESXi 8.0 covers everything the nmlx4_en driver ever served, and Broadcom’s own wording is blunt: “The driver does not exist in ESXi 8.0.” Not deprecated, not degraded — absent. NVIDIA reached the same place in its own stack earlier: ConnectX-3 and ConnectX-3 Pro stopped being supported as of MLNX_OFED 5.1, with anyone who still needs the card pointed at the 4.9 long-term-support branch. NVIDIA’s current ESXi driver line begins at ConnectX-4.

One window is still open, and it is worth knowing which one. The mlx4 module remains in the mainline Linux kernel, so a card already in a Proxmox or TrueNAS box will keep coming up. Leave a working one alone. Do not buy another. My earlier budget 10GbE build put its cheapest tier on these cards, and on a Linux host that tier still runs — on ESXi 8 it does not, and I would start any new build on something newer either way.

Used enterprise NICs are still the right buy

None of the above is an argument against used enterprise silicon. It is still excellent value, and the risk was never that the hardware dies. The risk is driver and firmware support, which is exactly what ended ConnectX-3.

So invert the shopping order. Check what your hypervisor’s current release ships a driver for, then look at prices. On ESXi that puts your floor at ConnectX-4. On Linux the window is wider, but “the kernel still builds the module” is a different thing from “someone is developing it,” and the gap between those two closes in one direction only.

What to buy

Most people reading this should not buy 10GbE. Buy storage.

If your NAS is one or two spinning disks, the cable is not what is slowing you down and a switch will not change your transfer rate by a meaningful amount. The same money in an NVMe drive, or in enough spindles to read in parallel, moves the number the network cannot. Turn on SMB Multichannel while you are at it — it is free and it is the second most common cause.

If your pool already clears a gigabyte a second, the shopping order is short:

  1. A fanless SFP+ switch. The CRS305 for four ports, the CRS309 for eight. If you buy a fanned switch anyway, budget for the replacement fan on the same order.
  2. Passive DAC for anything under five meters. Fiber past that, and only past that.
  3. ConnectX-4 or newer. Used, secondhand, from whoever is cheapest — but check the driver first and the price second.

The 2.5GbE comparison is the other answer here. For a NAS on one or two spinning disks, 2.5GbE is already faster than the drives, and it costs a fraction of this. That is not a consolation prize; for most homes it is the correct stopping point.

Wrong for

Wrong if you want silence without a fan swap. Buy passive-cooled or plan the replacement into the purchase — do not buy a fanned switch and hope.

Wrong if you expect 10GbE to fix a storage problem. It will move your bottleneck from a place you can see to a place you cannot, and your copies will finish at the same time.

Wrong if the card is too old for your hypervisor. The price on a dead-driver NIC is not a bargain, it is the market telling you something.

Frequently asked questions

Will adding 10GbE make my NAS faster?
Only if the storage behind it is already faster than the link you have now. Ten gigabits is 1.25 GB/s. A single Seagate Exos X24 is rated for 285 MB/s at its outer tracks, so one spinning disk hands you under a quarter of a 10GbE link and roughly twice a gigabit link. If your NAS is one or two hard drives, the cable was never the limit and a new switch will not move the number. Spend the money on NVMe or on more spindles reading in parallel, then revisit the network.
DAC or fiber for a 10GbE run at home?
Passive DAC for anything short, fiber past that, and Cisco’s own SFP+ datasheet puts a number on short: it lists passive Twinax at 1, 1.5, 2, 2.5, 3, 4 and 5 meters, with active Twinax starting at 7. DAC is cheaper, runs cooler, and draws almost nothing — 0.1W against 1W for an SFP-10G-SR module, and an optical link needs two of those. It is also one part instead of three, which matters more than the power saving when something stops linking at 2am.
Should I buy a used Mellanox ConnectX-3 in 2026?
No. VMware’s list of devices dropped in ESXi 8.0 covers everything the nmlx4_en driver served, and Broadcom’s own wording is that the driver does not exist in that release. NVIDIA stopped supporting ConnectX-3 in its own driver stack at MLNX_OFED 5.1 and points anyone who needs it at the 4.9 long-term-support branch, and the current ESXi driver line starts at ConnectX-4. The mlx4 module is still in the mainline Linux kernel, so a card you already own will still come up under Proxmox or TrueNAS. That is not a reason to buy one.
Why does a single file copy not saturate 10GbE?
Because one SMB session without Multichannel opens one TCP connection, and Microsoft’s own description is that the connection gets affinitized to a single CPU core. A faster switch does not give you a second core. SMB Multichannel is what fixes it: on an RSS-capable NIC it opens several connections for the same session and spreads them across cores. It is on by default on Windows. On TrueNAS SCALE it is a toggle under Advanced Settings in the SMB service, and it is worth turning on before you blame the network.

Evidence ledger

Last updated
Methodology
This networking guide was written and edited by Lowell K. Wood IV in St. Louis County, MO. Specs and prices verified against vendor and project documentation current on the date above. Full editorial standard: methodology. The fan-noise call, the DAC-versus-fiber call and the bottleneck call are mine — positions I hold from working with this class of gear, not bench results. Every number below is vendor or Microsoft documentation, cited inline; I measured none of it myself, and I have not published a rig inventory to back it.
Update log
  • 2026-09-07 — Last reviewed and updated.
Corrections
Spotted an error or a stale number? Email hello@techfuelhq.com. Confirmed corrections are added to the update log above.

About the author

Written by Lowell K. Wood IV, who builds and runs TechFuelHQ from St. Louis, Missouri.