If your Mac has pcie slots (i.e. not a Mac mini etc) then a regular Connectx-4 works out of the box. No tinkering.
ATTO 3102- Thunderbolt Mellanox ConnectX-4 Lx (Dual Port 25GbE) - $189 or less
- Thread starter nimik
- Start date
Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.
It works, however Atto's driver is slightly better. Actually in recent OSX Updates (I've noticed on 15.6 when I was doing some tests with aliexpress card) it seems Apple fixed CPU consumption and now it works as expected in that regard, but if you want jumbo frames you would need Atto's kext and matching PSID.
I found a guide on flashing. Haven’t read into it but you can check for yourself. Mellanox ConnectX-3 40 GbE using ATTO FastFrame macOS driver
I read on another forum that Chelsio T6 cards work fine with jumbo frames (9000). Might buy an additional to test.
Well, if you use their driver they should work, and same for CX4 - you need ATTO's driver (basically atto's firmware or modify stock one to change ids) and jumbo will work. That is just a limitation of built-in osx driver.
I still have one Chelsio T6225, but I don't like the card - it is way hotter than CX4 Lx and rather problematic on Linux (I haven't tried to put it into TB4 enclosure).
Ah, so with the CX4, you can use atto's drivers for it on mac and jumbo frames will just work? Just read the entire thread, and it looks like I can just cross flash the firmware and change the psid. Will try that when I get back. I also just purchased a T6225 to see. Anyways, now I have purchased too many network cards and have not even tested any one yet. I was originally thinking of using a 100gbe cx-4 to see but if the performance is worst than the t6225 or 25gbe cx-4, I will just move it to another server.
Last edited:
Sort-of. You still need to reflash CX4 Lx with atto's firmware or modify stock one.
In TB3/TB4/USB4 enclosures you are limited by roughly 32 Gbps total bandwidth. Co CX4 50G/100G even on Apple's driver might be tiny bit faster than 25 Gbps NIC.
I'm kinda waiting for affordable TB5 -> PCIe Enclosures to try CX5 PCIe Gen4 NIC, as that should give a proper boost to ~50 Gbps.
Yeah let's see... I ended up down the rabbithole and now I have an Atto on the way, 100 gb connectx-4 nic, 25 gb connectx-4 nic, and a 25gb chelsio nic...
For TB 5 enclosures, it's still so expensive to buy a pcie 4 100gb nic just to get a little boost. I was debating that too. A CX-5 dual port 100gb nic is at minimum $250 right now. So $350+250 = 600 probably can get you 50 gbps I guess then.
I also don't have any thunderbolt 5 devices yet unless I plan on getting a new macbook pro.
I finally got to work with the Mellanox ConnectX-4 Lx single-port and dual-port 25 Gb NICs. Picked them up with the OWC Mercury Helios 3s Enclosure (thanks @reasonsandreasons).
At first I hit a wall because I was flashing the wrong firmware — I mistakenly used fw_TLNS3252_32.1010.bin (for the Atto Thunderlink) instead of fw_FFN322_32.1010.bin (which is correct for the standard ConnectX-4 Lx). With the wrong firmware, there would be no link. Once I fixed that, the cards came up fine (I also flashed the dual port firmware onto the single-port ConnectX-4 Lx which worked) and I moved on to tuning Windows and macOS for better throughput.
Windows tuning
With these tweaks, single-stream tests went from inconsistent ~14–16 Gbps to sustained 20–22 Gbps, and with 8 streams (-P 8) I was able to max out around 23.3 Gbps (basically line-rate once you account for protocol overhead).
macOS tuning
macOS is trickier because the TCP stack has hard limits, but I was still able to improve stability a bit:
With these, Mac → Windows transfers averaged ~15–19 Gbps with occasional bursts to 22 Gbps. Windows → Mac was steadier, often in the 20–22 Gbps range. The hardware itself is fine (proved by multi-stream and UDP tests hitting full line-rate), but macOS just has tighter TCP buffer constraints than Windows or Linux.
Bottom line
At first I hit a wall because I was flashing the wrong firmware — I mistakenly used fw_TLNS3252_32.1010.bin (for the Atto Thunderlink) instead of fw_FFN322_32.1010.bin (which is correct for the standard ConnectX-4 Lx). With the wrong firmware, there would be no link. Once I fixed that, the cards came up fine (I also flashed the dual port firmware onto the single-port ConnectX-4 Lx which worked) and I moved on to tuning Windows and macOS for better throughput.
Windows tuning
- Enabled aggressive autotuning:
netsh int tcp set global autotuninglevel=experimental
→ immediately made single-stream speeds much faster. - Switched congestion control to CUBIC:
netsh int tcp set supplemental template=internet congestionprovider=cubic - Raised TCP memory limits via registry (64 MB send/receive window):
reg add "HKLM\SYSTEM\CurrentControlSet\Services\AFD\Parameters" /v DefaultSendWindow /t REG_DWORD /d 67108864 /f
reg add "HKLM\SYSTEM\CurrentControlSet\Services\AFD\Parameters" /v DefaultReceiveWindow /t REG_DWORD /d 67108864 /f - 9000 MTU Jumbo Frames
With these tweaks, single-stream tests went from inconsistent ~14–16 Gbps to sustained 20–22 Gbps, and with 8 streams (-P 8) I was able to max out around 23.3 Gbps (basically line-rate once you account for protocol overhead).
macOS tuning
macOS is trickier because the TCP stack has hard limits, but I was still able to improve stability a bit:
- Raised max socket buffer size:
sudo sysctl -w kern.ipc.maxsockbuf=16777216 - Increased send/receive space to 4 MB each (under the 16 MB cap):
sudo sysctl -w net.inet.tcp.sendspace=4194304
sudo sysctl -w net.inet.tcp.recvspace=4194304 - 9000 MTU Jumbo Frames
With these, Mac → Windows transfers averaged ~15–19 Gbps with occasional bursts to 22 Gbps. Windows → Mac was steadier, often in the 20–22 Gbps range. The hardware itself is fine (proved by multi-stream and UDP tests hitting full line-rate), but macOS just has tighter TCP buffer constraints than Windows or Linux.
Bottom line
- The CX4 Lx (both single port and dual port) work great in the Helios 3s once flashed with the correct firmware.
- Windows tuning (CUBIC + experimental autotuning + bigger buffers) unlocks full 25 GbE potential.
- macOS is the limiting side — still fast, but single-stream TCP is capped by kernel constraints. Multi-stream tests show the NIC and cabling can handle line-rate just fine.
- I still need to test the Chelsio 110-1209-60 T6225-CR 25/10GbE adapter which has a native driver to download on mac. It should end up being the same speed.
- I also want to test the 100 GbE Cx-4 but it will probably just throttle to around 24.8 Gbps.
- My Attos ThunderLink is coming too though it will probably have the same results as the Cx4 Lx.
- Still need to configure SMB, NFS to see what's throttling me.
Finally got my Chelsio 110-1209-60 T6225-CR card and it's even faster than my Connectx-4 card on my OWC Mercury Helios 3s Enclosure. It runs much hotter as even the transceiver is warm to touch and the card is pretty hot. I only installed the drivers and set jumbo frames. T6225-CR card -> my switch -> Connectx-4 card on my nas.
Last edited:
Regarding these AliExpress sfp28 units, has anyone figured out who made the carrier PCIe to Thunderbolt board? It seems to have a logo of an octopus inside a chip on the board. They seem to overheat and crash reliably, but running it outside of the case helps a lot. I've got an STL built to house it with better airflow and potentially a fan, but still need to test and rev it a bit.
The heat seems to be centered around the connectx card between U26 and U23, and the carrier board opposite the giant heat sink on the connectx. Because of course that was a great design decision...
The heat seems to be centered around the connectx card between U26 and U23, and the carrier board opposite the giant heat sink on the connectx. Because of course that was a great design decision...
Just got around the time to flash my Attos TLN3-3102-D00. I got it on eBay used but it was new and sealed. Had to flash it on my linux machine with a thunderbolt port as it's not possible on the mac. I'm getting around 19.7 Gbits/sec to my unraid machine. Unfortunately it doesn't look that good compared to my thunderbolt 3 enclosures maybe due to the thunderbolt chip it's using? I know the Sonnet uses a Titan Ridge controller. It's still fine since I don't think I'm pushing that much data often enough to matter. There is small benefit of using the Attos though. I don't think my Chelsio nic is renegotiating properly with my switch and I need to do it manually every time I unplug it I think. Will need test it more and see if it's just a Chelsio issue vs Mellanox.
After adjusting my SMB settings to have RSS and Multichannel on Unraid, I can finally max out my 25 Gbe connection to my macbook pro. 
Whoa - you've struck gold here! First time posting, myself. Wondering if you encountered the problem I'm having, and if you might have any tips? I'm getting read speeds like you are, but my writes to my NAS are about half of what you're getting at about 1300MB/s.
Setup: I've got the same T6225-CR card in an OWC Mercury Helios 3s Enclosure, connected to my M4 Max MacBook Pro, running the latest Chelsio v1.25.1 driver. In my NAS I've got a Mellanox ConnectX-5 card, running Rocky Linux with Houston UI. I've got a striped NVME in the NAS that can do north of 6GB/s. Some things I've tried / tested -
- iperf3 both directions are essentially line rate
- RSS and SMB Multichannel are enabled on the NAS
- jumbo frames enabled on both ends (though, interestingly, trying to run 'ping -D -s 8972 <IP ADDR>' fails Mac > NAS, but is successful NAS > Mac; I'm wondering if there is something about jumbo frames on the Chelsio end that isn't working)
- Attempted your macOS tunings from above with negligible benefits; those being:
- sudo sysctl -w kern.ipc.maxsockbuf=16777216
- sudo sysctl -w net.inet.tcp.sendspace=4194304
- sudo sysctl -w net.inet.tcp.recvspace=4194304
What speedtest app are you using? Also, with ping, you need to set sudo sysctl -w net.inet.raw.maxdgram=16384 for icmp to work. I would use AJA System Test Lite to get a more accurate reading of your speed. Is it a direct connection from your mac to your nas or through a switch? iperf3 is working at 22-23.7 Gbits/sec?
- Ah, didn't know about setting that sysctl setting for icmp to work. Will test later today but that probably solves that issue.
- Ya, I used AJA System Test Lite and Blackmagic Disk Speed test and just a copy using Finder; they all agree with those results - about 2.2GB/s read from the NAS but 1.3GB/s write.
- It's a direct connection with static IP's between Macbook and NAS; no switch.
- Yes, iperf3 is giving me around 24gbit/s both ways
Update: interesting, after running 'sudo sysctl -w net.inet.raw.maxdgram=16384' I still can't ping with MTU size of 8972 from the Mac.
Could you confirm that your qset info is any different from mine?
Could you confirm that your qset info is any different from mine?
FWIW, I've come up with some conclusions/changes that worked:
- Biggest thing was switching NAS OS to TrueNAS SCALE away from Rocky 9 / 45Drives Houston UI. That got me full iperf3 speed and file copy speeds to match - 3GB/s. (both of these OS's were running on top of Proxmox 9, but the 25Gbe NIC and HBA were passed through to both).
- Chelsio was able to reproduce the 'ping -D -s 8972 <IP ADDR>' so I'm waiting to hear back from them there, although it doesn't really seem to matter much.
I did find the individual firmware bins in the linux download zip. Currently searching for the atflash binary for MacOS!
Located the atflash executable under /Applications/ATTO FastFrame 3.app/Contents/Resources/ but it doesn't seem to like binary files. Will try using a linux or Windows box and the Mellanox tools to get this flashed!
Last edited: