A 10 Gigabit network sounds simple on paper: install a suitable network card, connect an SFP+ module or DAC cable, attach it to the switch, install the driver \u2013 done.<\/p>\n
In practice, however, the weak point is often not the server, the NAS, or the switch. The real problem can be the Windows client.<\/p>\n
In my case, the 10 Gigabit link basically worked. The switch was stable, TrueNAS was able to deliver 10 Gbit\/s, and the cabling was not the issue. Still, the Windows side showed instability, system freezes, and uncertainty around the network card being used. In the end, the solution was a different card, targeted cooling, and proper temperature monitoring.<\/p>\n
The goal was a fast internal network for large amounts of data:<\/p>\n
The core infrastructure consisted of:<\/p>\n
On paper, everything looked right. In reality, however, 10 Gbit\/s is not just about the raw link speed. PCIe slot behavior, drivers, the network card itself, temperature, airflow, and Windows behavior all matter.<\/p>\n
The first Intel SFP+ card caused problems in the Windows PC. There were system freezes and general instability. In another system or environment, the same type of card could behave much more normally.<\/p>\n
This is an important point: a network card can technically work and still cause problems in one specific desktop PC.<\/p>\n
Possible causes include:<\/p>\n
Used server network cards are often not designed with normal desktop cases in mind. They usually come from servers with strong, directed airflow. In a tower PC, that airflow is often missing.<\/p>\n
Many 10 Gigabit SFP+ cards are passively cooled. In a server rack, this is usually fine because multiple chassis fans constantly push air through the system.<\/p>\n
In a desktop PC, the situation is different. The card often sits below the graphics card, close to a PSU shroud, or next to other components. The small heatsink on the network card may receive some ambient airflow, but usually no targeted airflow.<\/p>\n
This can work \u2013 but it does not have to.<\/p>\n
The key lesson:<\/strong> with 10 Gigabit cards, do not guess the temperature. Measure it.<\/p>\n The card temperature is often not visible in Device Manager or common monitoring tools. The card can get hot without Windows clearly showing it.<\/p>\n With Mellanox\/NVIDIA cards, however, the temperature can be read using the Mellanox Firmware Tools. In my case, the card was detected with Example:<\/p>\n The output was a simple number, for example:<\/p>\n This means 43 \u00b0C.<\/p>\n The ASUS motherboard offered several sensor sources for fan control, including CPU, System, PCIEX16, and PCIEX8.<\/p>\n Important: these PCIEX sensors do not read the internal temperature of the Mellanox card. They are most likely motherboard sensors located near the PCIe slots. They can be useful as a rough ambient temperature reference, but they do not replace the real card temperature reported by They can still be useful for a fan curve, but it is important to understand what is actually being measured.<\/p>\n For cooling, a small 40 mm 12 V fan was mounted directly onto the heatsink of the Mellanox card. The fan only had two wires:<\/p>\n No tachometer signal. No PWM.<\/p>\n This still works on a normal ASUS CHA_FAN header if the header is set to DC mode in the BIOS. In DC mode, the motherboard controls the fan speed by changing the voltage. The downside: the motherboard does not know whether the fan is actually spinning. It will typically show 0 RPM or N\/A.<\/p>\n Therefore, the fan monitoring for this header has to be set to \u201cIgnore\u201d or \u201cLow Limit Ignore\u201d.<\/p>\n The Intel card was no longer treated as the only possible solution. Instead, a Mellanox\/ConnectX card was installed.<\/p>\n The advantage: solid server-grade hardware, a good driver base, and with the Mellanox Firmware Tools, a useful way to inspect the card\u2019s status.<\/p>\n The card was detected under Windows using MFT:<\/p>\n Example output:<\/p>\n After that, the temperature could be read:<\/p>\n A 40 mm fan was mounted on the existing heatsink of the card. The fan was screwed directly onto the heatsink so that air was pushed directly across the fins.<\/p>\n This is not an elegant manufacturer-designed solution, but it is a very effective practical fix.<\/p>\n Important points for this kind of modification:<\/p>\n The fan was powered from a motherboard fan header.<\/p>\n BIOS settings:<\/p>\n A Noctua Low-Noise Adapter was also used to reduce fan noise. These adapters essentially work by adding resistance, reducing the effective voltage\/current and therefore lowering fan speed. The important point is that the fan must still start reliably during a cold boot.<\/p>\n The card temperature was checked several times.<\/p>\n Open case:<\/p>\n Closed case:<\/p>\n This confirmed that the sensor was reporting plausible values. The temperature rose slightly when the case was closed, but remained stable in a very good range.<\/p>\n A longer iperf3 test was used to verify the connection. The link ran for several minutes at around:<\/p>\n After roughly 465 seconds, more than 500 GBytes had been transferred. The Mellanox card temperature remained around:<\/p>\n This is a very good result. With poor airflow, a stronger temperature increase would normally be expected after several minutes of continuous 10 Gigabit traffic.<\/p>\n Installing a 10 Gigabit network card is not enough. The whole system has to be considered:<\/p>\n Desktop PCs are not automatically ideal environments for passive server network cards.<\/p>\n Many SFP+ cards are passively cooled because they are designed for chassis with constant airflow. In a normal PC case, a small targeted fan can make a big difference.<\/p>\n A 40 mm fan directly on the heatsink is not beautiful, but it is effective. An even better final solution would be a 3D-printed air duct that guides airflow across the main heatsink and further over the SFP+ cage.<\/p>\n Possible improved final solution:<\/p>\n Touching the heatsink with a finger gives a rough impression, but it is not a measurement.<\/p>\n For Mellanox\/NVIDIA cards, This shows whether the card is really getting too hot or whether the cooling is sufficient.<\/p>\n File copies are only partially useful as network tests. They depend on many factors:<\/p>\n For a clean network test, iperf3 is much better. Only after iperf3 delivers stable 9\u201310 Gbit\/s does it make sense to further optimize SMB, NAS storage, and Windows file copies.<\/p>\n If iperf3 works properly but file copies do not reach the expected speed, the problem is probably no longer the physical 10 Gigabit connection.<\/p>\n Then the following areas should be checked:<\/p>\n For a stable 10 Gigabit setup on the Windows side, I would now proceed as follows:<\/p>\n After the modification, the result looked like this:<\/p>\n This confirmed that the Windows client side was stable, the card was sufficiently cooled, and the 10 Gigabit connection was technically working properly.<\/p>\n 10 Gbit\/s in a home network or home lab is absolutely achievable. But the Windows client side should not be underestimated.<\/p>\n The most important lesson from this setup: the problem was not simply \u201c10G does not work\u201d. It was a combination of hardware, drivers, desktop airflow, and missing temperature visibility.<\/p>\n With a stable Mellanox card, targeted cooling, and proper testing using iperf3, the connection runs reliably at almost full 10 Gigabit speed.<\/p>\n The real lesson: with 10 Gbit\/s, you have to work systematically. First stabilize the link and temperature. Then measure throughput. Only after that should SMB and storage performance be optimized.<\/p>\n Anyone who simply installs a used server card into a Windows PC and hopes everything will remain stable might get lucky. The better approach is: measure, cool, test \u2013 and only then optimize.<\/p>\n A 10 Gigabit network sounds simple on paper: install a suitable network card, connect an SFP+ module or DAC cable, attach it to the switch, install the driver \u2013 done. In practice, however, the weak point is often not the server, the NAS, or the switch. The real problem can be the Windows client. In […]<\/p>\n","protected":false},"author":1,"featured_media":1676,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[34],"tags":[],"class_list":["post-1674","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news-en"],"yoast_head":"\n3. Windows does not show the card temperature by default<\/h3>\n
mst status<\/code> and then checked using mget_temp<\/code>.<\/p>\nmst status\r\nmget_temp -d mt4117_pciconf0<\/code><\/pre>\n43<\/code><\/pre>\n4. BIOS fan sources do not read the Mellanox card directly<\/h3>\n
mget_temp<\/code>.<\/p>\n5. A 2-wire 12 V fan has no tachometer signal<\/h3>\n
\n
The Solution<\/h2>\n
1. Switching to a Mellanox\/ConnectX card<\/h3>\n
mst status<\/code><\/pre>\nMST devices:\r\n------------\r\nmt4117_pciconf0<\/code><\/pre>\nmget_temp -d mt4117_pciconf0<\/code><\/pre>\n2. Active cooling directly on the heatsink<\/h3>\n
\n
CHA_FAN\r\nMode: DC\r\nFan Speed Low Limit: Ignore\r\nSource: PCIEX16 or PCIEX8, depending on slot position<\/code><\/pre>\n3. Temperature testing with mget_temp<\/h3>\n
43 \u00b0C<\/code><\/pre>\n45 \u00b0C<\/code><\/pre>\n4. Load testing with iperf3<\/h3>\n
9.46\u20139.49 Gbit\/s<\/code><\/pre>\n45 \u00b0C<\/code><\/pre>\nWhat Can Be Learned from This?<\/h2>\n
1. 10 Gbit\/s needs more than just the right card<\/h3>\n
\n
2. Server hardware often needs server-style airflow<\/h3>\n
\n
3. Do not just touch the heatsink \u2013 measure the temperature<\/h3>\n
mget_temp<\/code> is very useful:<\/p>\nmget_temp -d mt4117_pciconf0<\/code><\/pre>\n4. iperf3 is mandatory<\/h3>\n
\n
5. SMB is a separate topic<\/h3>\n
\n
Recommended Approach for Stable 10 Gbit\/s on Windows<\/h2>\n
\n
mst status<\/code>.<\/li>\nmget_temp<\/code>.<\/li>\nExample: Stable Final State<\/h2>\n
\n
Conclusion<\/h2>\n
Further Reading<\/h2>\n
\n