Last month I reported some worrying results in the performance of some Thunderbolt 3 SSDs connected to Thunderbolt 4 hubs, where their write speeds could fall below 700 MB/s when writes were performed concurrently to other, faster storage. This article extends those with additional SSDs, and reveals circumstances where write speeds can fall as low as 400 MB/s.

Methods

In all now, I have tested five different Thunderbolt 3 SSDs:

• Disk A: 2 TB TB3 model claimed to achieve up to 2.8 GB/s and costing over $/€/£ 300.
• Disk B: 4 TB TB3/USB-C model claimed to achieve up to 2.8 GB/s and costing over $/€/£ 600.
• Disk C: 2 TB self-assembled TB3 model containing a Samsung 980 PRO M.2 NVMe SSD in Orico case K.
• Disk D: 2 TB self-assembled TB3 model containing a Samsung 990 PRO M.2 NVMe SSD in Orico case L, costing a little over $/€/£ 300.
• Disk E: 1 TB self-assembled TB3 model containing a Samsung 980 PRO M.2 NVMe SSD in Orico case L.

These additional tests reported here concentrate on combinations of Disks B, D and E, using the newer of the two Thunderbolt 4 hubs, with all cables being short, passive Thunderbolt 4 cables from CalDigit. Benchmarking software used was my own Stibium, for random read or write tests, consisting of 50 test files of random sizes in a uniform distribution from 2 MB to 2 GB, totalling over 45 GB in each test.

Results

These are summarised in the following updated table, which consolidates those from previous tests with this new series.

Much of this should be self-explanatory. For trio (hub) results, uniform refers to all three SSDs transferring data in the same direction at the same time, and mixed refers to one of the three reading or writing, while the other two are writing or reading, respectively. Trio (ports) refers to Disks D and E being connected direct to TB4 ports on the Mac Studio, with both those and the internal SSD performing the same operations at the same time.

When tested alone, all five SSDs attained speeds consistent with full four-lane Thunderbolt 3 performance, although when connected directly Disks A and B were slightly slower to write than C to E, and their write speeds fell to 1.5 GB/s when connected via the hub. Tested when one of a pair was performing either simultaneous reading or writing, there was little difference in their performance, even when connected via the hub. Total transfer rates remained close to 3 GB/s regardless of combination, and that total was shared fairly equally between each of the two SSDs.

Differences between Disks A and B on the one hand and C to E on the other only became pronounced when three SSDs were connected to the hub and were written to simultaneously. When two of those SSDs were Disks A and B, and the third was Disk C, write speed for A and B fell to about 700 MB/s. When only one of those SSDs was Disk A or B, and the other two were Disks D and E, write speed for A/B fell to about 400 MB/s. During that period, Disks C/D/E sustained write speeds of 1.6 GB/s, the same as when writing as a pair on the hub. Thus, Disks A and B suffered write throttling when competing for bandwidth against Disks C, D and E. Put in context, 400 MB/s is typical for an indifferent SATA SSD over USB 3.1 Gen 2, and less than 20% of their best write speed when connected direct to a port on the Mac.

Trio read speeds

When all three SSDs connected to the hub were reading, their speeds were a fairly uniform 1.1 (central range 1.0-1.2) GB/s, as shown in the chart below for the combination of Disks B (x points), D and E (open and filled circles).

In these tests, all three disks were reading within the first 2 seconds, and completed the test after 44 seconds had elapsed. For comparison, the graph below shows the same test performed with Disks A (*), B (+) and C (circles).

The period during which all three SSDs were reading simultaneously started after about 3 seconds had elapsed, and finished at about 47 seconds, when only Disk C was left completing its 100 reads, and returned to a read speed of about 2.6 GB/s.

These demonstrate that read bandwidth over the hub’s single TB3 connection to the Mac is evenly distributed across all three SSDs, regardless of which type they are.

Trio write speeds

When all three SSDs connected to the hub were writing, their speeds were very different. Speeds of Disks A or B fell to about 400 MB/s when they were writing at the same time as both Disks D and E, which sustained write speeds of 1.6 GB/s as they had when writing as a pair via the hub. This is shown clearly in the graph below, in which measurements for Disk B are shown as x, and Disks D and E as open and filled circles.

During the first 34 seconds, while Disks D and E were writing at about 1.6 GB/s, Disk B only wrote at about 400 MB/s. The latter continued at that low speed for a further 8 seconds after Disks D and E had completed their tests, before rising to a steady 1.5 GB/s, its maximum write speed when connected via the hub.

This compares with a similar phenomenon observed with the combination of Disks A, B and C during previous testing, and shown in the graph below. Write speed was preserved for the fastest of the three SSDs, Disk C (+), at 2.1 GB/s, but was throttled to about 700 MB/s on Disks A and B (* and o).

The time interval between the last write to Disk C and Disks A and B returning to a more normal write speed of about 1.4 GB/s was here about 10 seconds.

SSD type and performance

There are three types of Thunderbolt 3 SSD in common use:

• Full four-lane NVMe SSDs that deliver close to the maximum expected transfer speeds of 3 GB/s whether connected direct or via a hub, for both read and write. Bandwidth is shared across these equally, so that when more than one is connected to a hub, each gets an equal share, and their read and write speeds are predictable from the maximum available through the single TB4 connection to the Mac. Thus, two will read or write at about 1.6 GB/s each during simultaneous transfers in the same direction, and three at about 1.1 GB/s.
• Four-lane NVMe SSDs that perform similarly when reading, but whose write performance is throttled when connected via a hub. In the worst case, when competing with two faster SSDs on the same hub, their write speed can fall to 400 MB/s. Even when used alone on a hub, their write speed is limited to 1.5 GB/s.
• Two-lane NVMe SSDs that can only deliver read and write speeds up to about 1.5 GB/s even when connected direct to a Mac. I haven’t tested any of these, and don’t recommend their use where performance is important.

For consistent and best performance when used on hubs, only the first type is recommended. However, when a Thunderbolt 3 SSD isn’t going to be used via a hub, either of the first two types can be used.

These results also give insights into total Mac IO performance. Total read speed across the internal SSD and two TB3 disks (connected direct) was 10.2 GB/s, and total write speed was 11.9 GB/s. A single Thunderbolt port on the Mac Studio M1 Max is capable of transferring data both up and down simultaneously at a total rate of 4.7 GB/s, which exceeds the maximum unidirectional speed of about 3.2 GB/s, but falls well below twice that.

Recommendations

When selecting Thunderbolt 3 SSDs for use on a hub in situations in which write performance is important and simultaneous writes could happen to more than one attached SSD at a time, measure the write speed of each SSD when attached singly via the hub. If that’s below 1.6 GB/s, then that disk is likely to suffer write throttling via the hub. Although this might not appear a likely combination, the effects are serious: in this case, a costly SSD that delivered 2.2 GB/s connected direct to the Mac, reduced to 1.5 GB/s the moment it was connected to a hub, 0.7 GB/s when contending against one faster SSD, and 0.4 GB/s when up against two.

Quite why it does is open to speculation, but with the rising popularity of Thunderbolt 4 hubs, testers and reviewers of SSDs should measure performance when connected via a hub, as well as directly, to give prospective purchasers this important information.