Computer Parts: Marvell AQC113C 10GbE network card (QFly NIC-10G), PCIe x4 riser, and 7-port USB3 PCIe x1 Card.

Over the years, I've written about transitioning to faster ethernet here at home with standard copper 10GBASE-T (10GbE) cabling, beginning back in 2018. This works fine even with Cat-5e in the walls of my home (although some renovations use Cat-6). More recently, my home network was upgraded to 2.5GbE for any device that can benefit, with standard 1 gigabit/s fallback.

Last month, I needed another 10GbE network interface card (NIC) and noticed that we're now seeing inexpensive Marvell mGig AQC113C NICs available, upgrades from the AQC107 cards I have been using. So I grabbed a QFly NIC-10G (~US$75) to try out. There are other similar products like the NICGIGA (very unfortunate name), and GigaPlus.

The card offers 10G/5G/2.5/1GbE speeds and below (100/10Mbps) for compatibility. The new chip operates at only 4W running 10G speed (I believe the AQC107 is estimated at ~6W) making it even cooler-running for heavy loads.

As you can see in the picture of the card above, the interface is still PCIe x4 wide. Even though the AQC113C chip is capable of PCIe 4.0 which can do 16Gbps/lane bidirectionally (so even a x1 interface is fast enough), in order to be compatible with more common PCIe 3.0 (8Gbps/lane so will need at least x2 for full performance) motherboards and lower, it's a x4 card. I suspect in the future, we'll see x1 cards meant to run in PCIe 4.0 and latest 5.0 motherboards (32Gbps/lane!).

I was pleasantly surprised by the completeness of the QFly box considering the low price:

We have a low-profile bracket, extra screws, small pamphlet manual discussing driver installation, both a CD and USB stick with drivers, and even a screwdriver.                 

For me, I installed it in my Windows Server 2019 machine and it's compatible with the same driver I had been using for the ASUS AQC107 card (3.1.6 worked fine for Server 2019, but not the latest 3.1.8 which works well in Windows 11). Keep an eye on the Marvell website for newer driver/firmware versions and other OS like Mac and Linux.

Since my network switches are compatible with jumbo frames, I set my receive and transmit ring buffer size to 3000 each to keep it simple (26MB RAM x 2 for receive and transmit). Other settings like Energy-Efficient Ethernet, large send offload, checksum offload, adaptive interrupt moderation, left on by default:

Regarding buffer size, I've seen the recommendation for transmit buffer to be 2x receive buffer size and have used that in the past. I don't see a reason unless there's a difference in how the kernel manages the interrupts based on the in/outbound data. I suppose if you want lower latency especially when receiving data, lowering the receive buffer size could be useful. In general, larger buffer size reduces packet loss by making sure data doesn't overflow during the time between kernel interrupts (and reducing the number of interrupts decreases CPU load), but can increase latency for apps needing the data sitting in the buffer (in particular real-time applications like gaming).

So how fast is the card? Here's a current network map and the 1-hour iPerf speed test between the ASUS XG-C100C (AQC107 chip) in my Workstation and the new QFly (ACQ113C chip) in the Server computer. There are two switches along the path, the QNAP QSW-M2108 and Netgear GS110MX:

Using latest iPerf3.16 x64:
Workstation: "iperf3.exe -s -i 60"
Server: "iperf3.exe -c 192.168.1.10 -t 3600 -i 60 -l 128k"

Excellent. Average 9.88Gbps across 60 minutes running full tilt with 4TB transferred across the hour. I was doing some light blog editing in the first 30 minutes, there's a web server, Roon server (not playing music at the time) running in the background causing some of the fluctuations. Lowest minute average was still at 9.71Gbps while editing the blog.

Of course in real life it would be very unusual to have the network transferring data constantly like this over an hour! Despite the load, the QFly card's heat sink was barely warm to the touch at the end of the hour. I have confirmed speed stability with the same test settings over 8 hours between the ASUS XG-C100C and QFly NIC-10G - no sweat!

So far, after about a month of 24/7 use in the Server machine (Roon, LMS, website, files), the QFly AQC113C card continues to work well with no issues, excellent speed stability, running cool. Going forward I would transition from the AQC107 to these AQC113C if I need to buy new cards.

--------------------

I must admit that after having a 10Gbps LAN pipeline for a number of years, dropping down to even 2.5Gbps was noticeable with large file transfers. Once one experiences this kind of speed, it's hard to go back! So, recently when I added the massive nVidia RTX 4090 GPU to my Workstation, the darn thing covered my only other PCIe x4 slot so I couldn't get the 10GbE NIC in there. 🙁

BTW, here's a card size comparison between my old ASUS nVidia GTX 1080 vs. the Gigabyte RTX 4090 (for context, that square tile measures 18"/side):

In order to circumvent this, I needed to grab a PCIe extender so I could take the x4 signal from below the GPU and mount the NIC vertically in a card slot that my ATX full tower case conveniently had available.

Here's the PCIe extender I used (less than US$15), capable of PCIe 3.0 x4 speeds for the 10GbE card. As a PCIe 3.0 product, if you're connecting this to a PCIe 4.0 card, you might need to go into the BIOS to limit the slot to PCIe 3.0 speeds. For my needs, it was important that the male end didn't rise too much above the motherboard connector and the cable is flexible enough to clear the gap below the RTX 4090. Also important to check the length, 23cm was more than enough for me.

Notice the white 4-pin JST power connector if needed on the female PCIe end. Not needed for the low-power 10GbE network card.

Some extenders have 90° connectors that might work better for you.

Here it is all connected up in my chassis:

Friction is strong enough for the extender's connection with the ASUS card to remain in place.

With that, I'm back to 10GbE network speed on the Workstation, achieving the average 9.88Gbps transfer performance as shown above. Feeling good again...

Stable, fast network = Serenity.

As mainly an audiophile blog, permit me to remind audiophiles that ethernet is inherently galvanically isolated. Don't be impressed about snake oil ethernet switches for example or concerns about ethernet cabling (discussed close to a decade back!). Notice that many switches in the audiophile market are only 100Mbps devices likely because the companies are low-tech, silly, and want you to think that higher speed ones, even just 1Gbps (!) somehow are more "noisy" or "jittery" and can somehow affect your DAC's performance. That's as usual just rubbish unless they can prove such a thing and show in what context they find these issues.

Buyer beware for any products intended to create an "audiophile network" - whatever that is. 🤔

--------------------

One last item to mention. Here's an inexpensive PCIe 1x single-card USB 3.0 interface (less than US$30). I bought one to add a few more USB ports on my gaming PC for the VR system.

The 7 ports include 4 USB-A and 3 USB-C connectors should be plenty, and it's convenient not needing to take external power off the power supply like a Molex connector or the SATA power cable (like some cards). The company claims up to 5W available over the USB-A ports, and 15W for the USB-C. Windows 11 already has native drivers for the Fresco Logic FL1100 controller (paired with VLI VL817 hub), also supported under Linux and MacOS. The USB ports, arranged perpendicularly along the card worked fine.

I know some users have experienced USB disconnection issues with AMD 5000 CPUs and certain chipsets like their B550/X570 boards under heavy processing loads with PCIe 4.0 graphics cards. Some folks running AMD systems still complain of USB disconnection issues even after BIOS updates and chipset driver updates. Workarounds like putting the PCIe x16 down to Gen 3, disabling C-States, or playing around with the NVMe M.2 configuration also helps but of course this should not be necessary. Anyways, this USB card is cheap, works well for 5Gbps speeds (PCIe 2.0 1x fast enough), I suspect would be a good solution for those who still have issues and can benefit from extra ports.

There is a similar USB 3.2/3.1 Gen 2 version of this card (around US$45) capable of 10Gbps if you have a PCIe x4 slot free.

As with the mythical "audiophile network", I don't see a reason why one needs "audiophile USB ports" since Bits Are Bits and high quality DACs in objective testing maintain low noise/jitter regardless of USB source. Would love to see objective evidence that an expensive €850 JCAT NIC can even make a difference in DAC performance!

--------------------

Happy high-speed computing and home networking, everyone! Amazing that it's already February.