Published May 13, 2021

The M1 Mac mini drew wide praise for its performance and price. It outperformed every previous Intel Mac in single-threaded benchmarks and a Ryzen 5 5600X, one of the fastest single-thread desktop CPUs on the market, at a lower cost. Its multithreaded performance trailed its single-threaded results because it has fewer cores.

Input and output (I/O) has drawn more criticism. The base model includes two USB ports, two Thunderbolt ports, and one gigabit Ethernet port. Users who needed faster networking had two options: a bulky, expensive dongle that occupied a Thunderbolt port, or an Intel platform. Apple had previously offered 10-gigabit Ethernet on the iMac Pro and on Intel-based Mac minis.

This article examines a 10-gigabit M1 Mac mini tested with Iperf in server mode. The tests measured network throughput under increasing load from displays, Thunderbolt and USB storage, and other attached devices, to identify internal bandwidth limits.

Test Objectives

The tests addressed two questions:

1. Whether Apple delayed the 10-gigabit version of the M1 Mac mini because of performance concerns.
2. How the M1 Mac mini allocates limited I/O bandwidth across its ports and internal components.

Apple does not publish block diagrams of its platforms, as Intel does. Internal bottlenecks therefore appear only as performance drops during use.

Baseline Network Performance

The first test ran with only network and power cables connected, with no display attached.

The 10-gigabit line rate is the signaling speed. Actual data throughput is normally lower.

Display Connection

The second test connected a Pro Display XDR. The M1 Mac mini is the first Mac mini to support Display Stream Compression (DSC). DSC leaves enough bandwidth to run the monitor's integrated USB 3 hub at full speed.

Earlier Intel-based Mac minis either did not support the monitor's full 6K resolution or drove it in an uncompressed tiled mode. Some users reported that this mode reduced the USB hub to USB 2 speeds. Uncompressed, the Pro Display XDR requires about 34 Gbps of bandwidth.

Network throughput did not change with the display connected. The test illustrated that the system divides a finite pool of bandwidth among its components.

Thunderbolt Storage Load

The third test added a high-performance Thunderbolt SSD. Anthony adapted a shell script that writes data continuously to an SSD and logs speed at regular intervals.

• The SSD read speed exceeded 10 Gbps.
• Apple advertises up to 40 Gbps for each Thunderbolt port.
• The two Thunderbolt ports did not share bandwidth in this test.

USB-C Storage on the Display Hub

The fourth test connected three USB-C SSDs to the USB hub of the Pro Display XDR and ran the storage script on all three simultaneously.

The SSDs slowed under load, which identified a bottleneck in the display's USB hub. This result did not reveal any internal bottleneck of the M1 Mac mini.

Expanded Load Test

The final load test used five SSDs on the display hub and one additional Thunderbolt 3 SSD on the Thunderbolt daisy chain.

One drive did not appear correctly. Running the script against a missing target increased CPU load. After the drive was recognized, network throughput ranged from 9.3 to 9.42 Gbps with all drives active.

CPU usage rose as additional storage activity was added:

The script's CPU overhead affected results but did not change the network results. Runaway CPU usage does reduce network transfer performance.

Key Findings

The 10-Gigabit Port Does Not Share Bandwidth

The 10-gigabit Ethernet controller has a dedicated PCI Express Gen 4 lane. It has no switch chip or controller hub in its path. Testing found no bottleneck affecting this port. The test results indicate that Apple's delay was not caused by performance limits.

Storage Throughput Is Limited

• Theoretical maximum throughput for a single drive tested individually is just over 58 Gbps.
• Measured real-world throughput peaked near 50 Gbps.
• Throttling therefore occurred.

The SSDs connected to the Pro Display XDR were limited by its DisplayPort signal, which uses about 13.7 Gbps.

Each Thunderbolt 3 Port Caps at About 20 to 22 Gbps

Each Thunderbolt 3 port reached a ceiling of about 20 to 22 Gbps. The following factors account for the gap between the advertised 40 Gbps and the measured limit:

• About 7.6 Gbps is reserved for DisplayPort on the Thunderbolt 3 dock.
• Signal encoding consumes additional bandwidth.
• The resulting theoretical maximum is about 26 Gbps.

The measured results are close to this figure. The remaining difference is attributed to traffic through busy hubs rather than direct connections. This explains the wavy patterns on the graphs, compared with the steadier results from individual drives.

Recommendations for Latency-Sensitive Work

Applications sensitive to I/O variation, such as audio or real-time data workflows, perform best when devices attach directly to the computer. Where a hub or dock is unavoidable, only one high-speed device should be attached at a time.

For most users, these limits will not affect everyday work.

I/O Architecture

The test results indicate that the M1 Mac mini has four PCI Express lanes allocated to its I/O:

• Two PCI Express Gen 4 lanes connect to Apple's Thunderbolt controller, which provides two Thunderbolt 3 channels.
• One Gen 4 lane serves the 10-gigabit Ethernet controller.
• About 10 Gbps serves the Type-A USB ports. This allocation may share a lane with the Wi-Fi 6 adapter.
• The HDMI port connects directly to the GPU through an internal DisplayPort-to-HDMI converter.

The macOS terminal command ioreg displays device hierarchies but does not show how devices connect to the main system.

Storage Architecture

Most PC systems provide more than four PCI Express lanes. The M1 does not provide additional lanes for its SSD.

The T2 chip, which served as the storage controller in the final generation of Intel Macs, does not exist in the M1 generation. Reverse-engineering work to run Linux on the M1 indicates that the M1 system-on-chip (SoC) includes the security functions of the former T2 chip. This work also found that the SoC controls the SSD directly through a memory-mapped arrangement rather than through PCI Express.

NVMe is a protocol. It is commonly associated with PCI Express but does not require it.

Bandwidth for the internal GPU has not yet been accounted for in these tests.

Conclusions

• The M1 Mac mini is probably near the I/O limits of the M1 SoC.
• The M1 Mac mini has more I/O capacity than the initial analysis indicated.
• The reason Apple delayed the 10-gigabit model is likely related to logistics or strategy, according to the test authors. The tests did not establish this.