Aquantia/Marvell AQC107 10GbE Network Cards (eg. ASUS XG-C1001C): Driver and Firmware updates and improving stability. (And more audiophile ethernet snake oil warnings.) [Update: Reduce latency for Roon audio playback.]

While this blog is primarily about audiophile stuff, over the years I have discussed modern audiophile systems including thoughts on networks. Like it or not, computer audio/video is ubiquitous and probably has become the primary source of access for many of us these days (rather than deal with physical media). Furthermore, a fast, responsive, reliable home network is almost essential as a foundational "utility" for entertainment and work. As discussed back in 2018, I've been running 10 gigabit/s home ethernet (standard copper RJ45 Cat-5e/6) with an update to the ASUS router and QNAP switch infrastructure in 2021.

In 2022, when it comes to very high-speed computer network cards, the Aquantia/Marvell AQtion AQC107 chipset devices remain attractive due to reasonable price, PCIe 3.0 x4 compatibility, and ability to run at 2.5/5 gigabit intermediate speeds (also dropping down to 1Gbps and 100Mbps as needed of course). Over the years, I've seen comments online with frustrations around the popular cards like my ASUS XG-C1001C shown above. The issues are typically described as episodic disconnects usually for about 5 seconds (discussed here, here) and then coming back online. As you can imagine, this will lead to a time-out of network transmissions.

Depending on the application, this might or might not lead to problems. For audio and videophiles, this is a problem when we're streaming video to something like an AppleTV or playing music off Roon in real-time. Other apps may be more tolerant and can just wait a little longer and retry.

Various users have reported the cause of troubles originating from all kinds of issues like RF interference, to power line disruptions, to network cable issues. Since we are talking about very fast 10 gigabit/s transmissions, I suspect all these things could be to blame. While I've generally been fortunate, I have noticed these occasional disconnection issues showing up on my Windows Server 2019 computer as warnings a few times over 24 hours, seemingly randomly.

Example of AQC107 "network link lost" where it says "Warning aqnic640" shown in Windows Server 2019 events log. Often it happens in a cluster like this around 6:45PM randomly. Notice the previous disruption was four hours earlier at 2:43PM.

So for those who are experiencing this problem, here's what I did for both my Windows 11 Workstation and Windows Server 2019 machines I'm running ASUS XG-C100C cards on:

1. Update the NIC to the latest Marvell software: 1.8.0_3.1.121a firmware and 3.1.6 driver in Windows. Go here and search in "Marvell Public Drivers", "Windows" 64-bits only, and "AQC107" part number.

For Windows Server 2019, the latest 3.1.6 drivers installed but for some reason could not identify the card, so I went on Station-Drivers and got a slightly older 3.1.4 WHQL, installing the Windows 8.1 version works with the newer firmware under Server 2019.

2. Once you've done the above, go into the Device Manager and change a few of the Advanced settings including turning off power savings for the NIC like so:

Note that this is assuming you already know your network functions well at 10Gbps (ie. you have no issues 99.99% of the time!) and all you want is extra stability. These settings will force the NIC to only connect at 10Gbps, maintain the connection at full duplex, thus forcing auto negotiation to turn off. Turn off the Energy-Efficient Ethernet feature which is known to cause issues with some IP communications like Dante. Also, the "Recv Segment Coalescing" feature has been noted by folks to lead to episodic disconnects (I turn off both the IPv4 and IPv6 options), it also doesn't improve transfer speed in my system nor decreased CPU load.

If your network cabling is older or know you have inevitable issues at 10GbE, feel free to limit the link speed to 5 or 2.5GbE full duplex for stability.

Here's a rundown of some other settings in the "Advanced" tab:

Downshift retries: Disabled
Energy-Efficient Ethernet: Disabled
Flow Control: Disabled (usually unneeded overhead, and controversial benefit)
Interrupt Moderation: Enabled
Interrupt Moderation Rate: Adaptive
IPv4 Checksum Offload: Rx & Tx Enabled
Jumbo Packet: Disabled (make sure all devices compatible before use!)
Large Send Offload (all IPv4 and IPv6 options): Enabled
Receive Buffers: 2000 (max 4096 in multiples of 8)

Maximum number of RSS Queues: 8 (correlate to number of CPU cores)
Receive Side Scaling: Enabled
Recv Segment Coalescing (IPv4 and IPv6): Disabled
TCP/UDP Checksum Offload (IPv4 and IPv6): Rx & Tx Enabled
Transmit Buffers: 4000 (max 8184 in multiples of 8)

Log Link State Event: Disabled

[Regarding the buffer sizes and Jumbo Packets - see Addenda below if you're using Roon!]

Obviously some of the settings like receive/transmit buffers can be optimized for your situation; I've noticed best speed using the middle buffer sizes on my machines rather than max'ing out. "Interrupt Moderation" reduces the interrupt request frequency especially with high-speed networking, improving CPU efficiency. Feel free to adjust to "Low" or even "Disabled" if you need lower network latency (doing this along with lowering buffer sizes might be useful for gaming, will not affect audio streaming quality).

If you have a managed switch, you can also set it to lock at the same speed (ie. 10Gbps) as well. This is what I've done with my QNAP QSW-M2108-2C (port 10 is my Windows Server, port 9 connects to the upstairs Workstation):

Lock switch speed to 10GbE full duplex.

3. Now run some torture tests to see if stability is maintained (see below for more).

4. Still running into issues? Consider if the NIC could be overheating.

I noticed one of my two ASUS cards seemed to disconnect much more often than the other; the firmware, driver, and settings updates above did not stop the problem. If you still have issues especially if seeing disconnects during heavy data transfers on warm days, maybe it's heat. The AQC107 chip is small and notice that these passively cooled cards have relatively large heatsinks and during use, they warm up quite a bit.

So I took the problematic card out for an examination. I noticed that the thermal interface doesn't seem all that robust with the ASUS XG-C1001C. We can certainly improve it by getting our hands a little dirty. Let's take the heatsink off (easily done, remove the 4 small screws) for a look:

Notice that all we have here is a soft, thick thermal pad over the chip surface (1.2cm x 1.4cm). Remember a major lesson from Computer Repair 101 - "Only use a thin layer of thermal compound!" Let's scrape that stuff off and use some alcohol for a clean look at the chip:

Aquantia/Marvell AQC107 shiny and clean...

Notice the space where the thick thermal pad was. I really wonder how well heat was being transferred across the material and whether over time, the conduction ability may have deteriorated. Well, that obviously won't do! Let's get some physical metal contact for decent conductivity.

Alright then, I cut up a pure copper 0.5mm (24 gauge) sheet - about 2cm x 2cm will cover the chip but not intrude into the other nearby components:

Got some non-electrically conducting CoolerMaster CPU thermal compound ready. Make sure you use only a thin layer on both sides of the copper sheet.

Here's a side look with proper heat transfer interface.

Copper plate used as spacer to improve heat conduction/spread. Make sure the electrically conductive copper sheet isn't touching other components.

Now that's more like it! Good physical contact, make sure not to overtighten the four small screws keeping the heatsink in place. It has been said that the AQC107 running at 5Gbps uses 3W, likely pushing something like 6W running at 10Gbps. Even with the improved heat conduction, because this is a passively cooled card, make sure your computer has reasonably good airflow inside (maybe install the card close to a fan, or even strap on a 40mm fan).

To test, we can copy data across the 10GbE network and should see stable, high-speed transfers (477MB/s limited by SSD speed across SAMBA network):

Real-life 50GB directory copy across 10GbE.

Or use iPerf 3 and have a look at maximum transfer speed from the Windows Server 2019 to my Windows 11 Workstation across QNAP and Netgear switches over 7200 seconds (2 hours):

Performance test across network with managed QNAP QSW-M2108-2C and unmanaged Netgear GS110MX switches in the path. Windows 11 Workstation: iperf3.exe -s -i 60. Windows Server 2019: iperf3.exe -c 192.168.1.10 -t 7200 -P4 -i 60 -l 128k. (192.168.1.10 is the static address I keep my Workstation computer at.) See Addendum on update with newer version iPerf3, single stream without "-P4" option.

Excellent, almost 9.5Gbps average sustained speed with less than 10% CPU (Intel i7-7700K) utilization while running the network at full throttle over 2 hours while streaming music over Roon, web server running in the background, and some blog editing. Since I don't foresee ever running the network at maximum speed for 2 hours straight, this seems to be a reasonable "torture test".

As you can see with the last few minutes of iPerf data, the transfer rate stayed consistently above 9.4Gbps. Assuming you're not doing a lot of background data transfer, if you see the average value bouncing around especially dipping below 9Gbps, this could be a sign of overheating or poor cable quality. 

Also, while running the iPerf test, I can look at my QNAP switch port stats to see if it detects errors:

Errors do show up over time, I cleared the data before running the iPerf test.

Nice, at least at the level of this switch (again, 10GbE port 9 heading upstairs to my Workstation, port 10 connects directly to the Server), after more than 7.25 trillion bytes transmitted between the two computers, zero CRC and FCS (Frame Check Sequence) issues detected. And this is at almost 10 times the speed of the typical 1 gigabit ethernet found in most homes these days.

Hopefully this combination of firmware/driver update, hardware hack, and setting changes helps if you're also having issues. The Aquantia AQC107 cards are inexpensive 10GbE NICs so while I don't expect them to be as stable as enterprise solutions, I think they can work reliably in the home environment. At least with the ASUS XG-C1001C, the marginal thermal/heatsink interface can be improved.

Addendum:

Recently, I had a need for one more 10GbE card and got the TP-Link TX401 (~US$100), also based on the AQC107 chip. Note the slightly longer heatsink compared to the ASUS:

Cards are about same height with ASUS a little wider, TP-Link heat sink longer and probably radiates heat better with the fins.

Nice that TP-Link has included the half-height bracket and good 1.5m length of Cat-6A STP ethernet cable.

I updated the firmware and drivers to the Marvell OEM as above. Speed-wise it measured the same as the ASUS and out of the box, seemed quite stable already running at 10GbE for a few days now. The TP-Link NIC is named generically in Device Manager as "Marvell AQtion 10Gbit Network Adapter" and I see it has an extra Aquantia diagnostics panel which I had not seen before with the ASUS card:

 

Hmmm. Curious if anyone has seen this diagnostics panel with the ASUS AQC107 card or if you know how to activate this.

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

"DALL-E, show me a photo of an audiophile ethernet switch with a glass of snake oil sitting to the right side, on a table with a blue background."

Pretty cool that the algorithm drew me an ethernet switch with LED indicator (sound setting "C"?) and even a VU meter to indicate the peak awesomeness of the audiophile data passing thru. Maybe this switch can be used to "authenticate" the sound. ;-)

A final note to audiophiles.

Yes, networks can be complex systems especially when we mix-and-match various brands and multiple devices. Like it or not, modern technologically-based hobbies demand that we understand a thing or two about how to set things up properly and fix problems when they arise. Make sure to understand the technology enough so you don't get caught up in fear, uncertainty, and doubt; unfortunate factors IMO driving some of the beliefs and desires of consumers in this hobby encouraged by unscrupulous companies. (Unfortunately, this potential for certain consumers to become technologically embarrassed thus giving in to FUD can be seen on display in many audiophile reviews in magazines and online.)

It's not difficult to get networks functioning; products come with manuals to follow, and these days, there are a number of forums where one can get good technical help! However, the audio Industry continues to suggest that thousands of dollars spent on ridiculous ethernet cables could make a difference, and some still tout the inexplicable benefits of basically marked up ethernet switches from small companies like these (of course, I still would much prefer the NetGear Nighthawk S8000 if one can find one :-). And some still claim that otherwise bit-perfect, ethernet packet-based asynchronous data transmission can result in different sound qualities (yet they never can provide evidence for such beliefs).

It doesn't take much of a search to find examples such as this shilling of the Synergistic Ethernet Switch UEF (US$2300, only 5 ports) to improve the sound with "proper harmonic structure and tone". Obviously, this kind of review makes no sense (here's an even worse review of the Synergistic stuff). There's not even any reasonable putative mechanism for such a sonic change being offered by these people. Looking at these reviews and the Synergistic website, I can't even figure out something as basic as what speed the ethernet switch operates at! Can it even handle 1 gigabit speed?

Speaking of speed, it's amazing that many of these companies are only offering 100Mbps devices like the JCAT M12 Switch Magic for a mere €2550. Looks to me like they're just milking higher margins on old technology while hyping the novelty of using atypical M12 connectors ("used in railway in Japan", and factory automation - so what?)!

Not surprisingly, the moment anything is described as "audiophile", prices increase at least 100%. Stuff like the JCAT Net Card XE (€800) is based on the Intel i350 gigabit chip released in 2014, with unsubstantiated benefits of linear power upgrades and OCXO clocking (BTW, you can get a dual-port i350 NIC for about US$100). Or for the even more esoteric computer audiophile, how about the M.2-interfaced SOtM sNI-1G which is a relative steal at US$350 but increases to over US$1,000 with the "sCLK-EX" clock upgrade.

Regarding that clock upgrade for a grand, seriously folks, what are we doing this for with an ethernet card where the signal is often sent tens (even hundreds) of feet from the server computer to your streamer where the data will be reassembled, processed, and then more than likely fed to the DAC asynchronously which has its own clock!? Even if the ethernet card or downstream switch has "femtosecond" levels of accuracy internally, this would have no meaningful effect on the DAC at the end unless you're saying that the clocking issue is so severe that it caused data errors!

Furthermore, what kind of noise are they worried would be coming out of the NIC and into the ethernet cable that would affect the audio streamer, and even affect the DAC (let's not even worry about the switch(es) that could be along the path)!? Please, let's have a look at the evidence. Are we sure those who hype this stuff honestly believe what's being sold? Could they be delusional? Or could some of them simply be dishonest scammers, consciously shilling for the companies? (Gasp 

😱

!)

It would be nice to see some "truth in advertising" regulation governing companies like these in the audiophile cottage industry. But even if there were regulations, enforcement would be doubtful. The use of purely subjective opinions as "evidence" in audiophile magazines or online reviews have totally damaged the credibility of so much of the audiophile literature. Given the minefield of questionable hype and more-than-likely false claims, one really needs to tread carefully to determine what is valuable, and what must be ignored. Good luck audiophiles.

Speaking of "evidence", I see some companies like the Silent Angel folks have now included measurements like eye diagrams in their product descriptions like this for the Bonn N8 switch (8-port, 1GbE, asking price US$550). Does anyone out there put faith into that 369ps vs. 705ps jitter difference at the level of the ethernet switch having an effect on the sound quality coming from the downstream DAC?! Do not mistake these numbers for DAC jitter values; not the same thing and no correlation. Even that "power source signal" comparison presumably from the 5V/2A Forester F1 (US$600) power supply vs. maybe a generic switching wall-wart doesn't imply that the ethernet signal would be superior in any way using the expensive PS. While I appreciate that these Silent Angel boxes might be nicer looking, more reliable or better able to handle suboptimal network conditions compared to a much less expensive device, there's nothing here to suggest they "sound better" compared to any other bit-perfect ethernet switch. No, it's not true that "everything matters" when it comes to sound quality, and likewise not all measurable differences matter. The goal does involve gaining enough understanding, relevant experience, and wisdom to appraise the claims made so as to cut through rampant frivolous consumerism and pseudointellect in the audiophile hobby space.

[Don't forget that these days, high quality audio streaming through Logitech Media Server, Roon/RAAT, even UPnP/DLNA use TCP error-correcting data transfers which further ensure bit-perfection.]

Hope you're enjoying the music, dear audiophiles, as we enter October...

Addendum:

Thanks Etienne in the comments for the link to the new bug-fixed iPerf builds. Here's using the latest 64-bit version 3.12, single stream "iperf3.exe -c 192.168.1.10 -t 7200 -i 60 -l 128k" on my system:

Receive buffers = 2000, Transmit buffers = 4000. As per settings above.

Another addendum:

Turned on jumbo frames (9014 bytes setting in the NIC) for the Workstation and Server, switches support larger MTU. Doing this reduces the packet overhead, improving overall transmission speed, and reduced CPU cycles as well:

Receive buffers = 2000, Transmit buffers = 4000.

Not bad. About 0.4Gbps increase compared to standard 1500 byte payload with large transfers.