Achieving 25 Gbps Thunderbolt Ethernet on a Mac Studio

For several years, I have relied on the integrated 10 Gigabit Ethernet port on my Mac Studio. While it has been sufficient—allowing me to perform backups at roughly 1 GB/sec and edit 4K video directly from my NAS—I recently upgraded my network rack and NAS to 25 GbE. Naturally, I wanted my primary workstation to keep pace.

The Cost Barrier

When searching for 25G networking for macOS, I discovered that the market is prohibitively expensive. Because modern Macs lack internal PCIe slots (RIP Mac Pro), you are forced to use Thunderbolt adapters.

While the Raiden Digit is reasonably priced, I found a much cheaper alternative via Christian Kohlschütter: a Thunderbolt 3 adapter board paired with a server-pulled OCP 2 NIC.

Initially, this setup was a steal at $160. While Amazon prices have since climbed to$299, you can still find non-marked-up versions if you are willing to browse various Chinese marketplaces.

---

The Implementation Process

To get this running, I had to move beyond Cat6A and pull actual fiber to my desk. My journey followed this general workflow:

Solving the Software Bottleneck

Upon initial testing with iperf3 between the Mac and the NAS, I hit a wall. The NAS was running an outdated version of the tool.

The Fix: I compiled the most recent version of iperf3. This allowed the NAS to utilize multiple CPU cores for the transfer, pushing speeds to 20 Gbps. As noted by Christian, the Thunderbolt 3 chipset limits single-direction traffic to $\approx 20 \text{ Gbps}$ and bidirectional traffic to $\approx 25 \text{ Gbps}$, regardless of whether you use a Thunderbolt 5 port.

The Thermal Crisis

The hardware presented a more dangerous problem: extreme heat. The enclosure lacked a thermal bond to the OCP 2 network card. Since OCP 2 NICs are designed for high-pressure server environments, the tiny onboard heatsinks were useless in a passive box. The chips were essentially cooking.

Cooling Iterations:

• Phase 1: Added giant heatsinks to the exterior of the enclosure (Christian's method). Result: Stable, but still very hot.
• Phase 2: Installed low-profile internal heatsinks and a speed-controlled USB fan. Result: Too loud and insufficient airflow due to the enclosure plates.
• Phase 3: Full custom engineering.

The Final Engineering Solution

Thanks to some complimentary Prusament PLA in Noctua Brown and Beige, I decided to build a professional-grade cooling solution.

Components used:

• Fan: Noctua NF-A8 80mm 5V
• Controller: NA-FC1 Speed Controller
• Custom Parts: 3D printed fan duct and an airflow-optimized grill.

Technical Integration

I completely removed the front plate of the NIC enclosure to eliminate airflow resistance. I then spliced the Noctua braided extension cable and soldered it directly into the through-holes on the Thunderbolt-to-OCP PCB to source power.

The math for the power draw is as follows: $P_{total} = P_{NIC} + P_{fan}$ $P_{total} \approx 4.5\text{W} + 0.5\text{W} = 5.0\text{W}$

The voltage provided was approximately $4.8\text{V}$, which is close enough to $5\text{V}$ to operate the fan without risking a brownout for the NIC. I used Kapton tape for strain relief on the internal wiring.

Final Results & Performance

After assembly, the results were night and day. With the fan on a low setting, the temperature stabilized at under 36°C after 10 minutes of use, and the noise was virtually undetectable.

Performance Metrics:

• Network Throughput: Maxes at 20-25 Gbps (TB3 limitation).
• Samba Read Speed: $\approx 1.4 \text{ GB/sec}$
• Samba Write Speed: $\approx 1.0 \text{ GB/sec}$

The Verdict: These speeds are only marginally faster than the built-in 10G Ethernet. Given that I am using an Arm-based NAS (Ampere Altra) and SMB multichannel, the bottleneck likely lies in the CPU or protocol rather than the NVMe SSD array.

Was the effort of pulling fiber, 3D printing a cowling, and soldering a fan worth it for a slight bump in real-world speed? Probably not. But as a project, it was a success.

---

Related Reading:

• Remote power-on for Macs is finally possible.
• Achieving 1.5 TB of VRAM on Mac Studio.
• Exploring RDMA over Thunderbolt 5.
• The mystery of why Apple won't make a standalone Touch ID sensor.