When you invest in Thunderbolt 3 and 4 peripherals, you expect to you be rewarded with the best performance. As I suggested a month ago, fast high quality TB3 SSDs can slow down when connected to fast high quality TB4 hubs. This article extends those previous observations and confirms what you can expect.
This isn’t a review. The number of TB4 hubs that I’ve tested is only two, and I’ve only tested those with three different SSDs. I can’t tell you whether this is common, or confined to only those hubs and SSDs. However, looking at the few reviews of TB4 hubs that have reported performance, I’m not the first to experience this, although it appears to have been glossed over previously.
In case you haven’t realised it yet, the promise of 40 Gb/s over TB3 and TB4 isn’t all it might appear to be. In practice, only 32 Gb/s of the total bandwidth is available for data, including transfers with SSDs, as that’s the most Thunderbolt can get from its four lanes of PCI Express 3.0. Allowing for overhead, that means the fastest read and write speeds you can expect from an SSD over TB3 or TB4 amount to 3 GB/s.
As all devices connected to a hub share a single TB4 port on your Mac, the maximum total transfer speed of all data devices on that hub cannot exceed 3 GB/s. So if you had two fast SSDs both transferring data for writing, you might see them attaining a write speed of 1.5 GB/s each, but unable to exceed that. However, a single fast SSD should still approach its maximum of nearly 3 GB/s. What I show here demonstrates that, for some SSDs at least, their maximum write speed can be capped to half of that, or 1.5 GB/s, even when there’s no other device connected.
Currently, you’ll come across two main types of TB3 SSD: those capable of using all four lanes of PCIe 3.0, and approaching speeds of 3 GB/s, and those only able to use two lanes, so limited to 1.5 GB/s. The latter are generally older, less expensive, and don’t require the same cooling.
I have now used two M1 Macs as hosts for these tests, although these latest results were all obtained from my Mac Studio M1 Max running Ventura 13.2.1. Two different TB4 hubs were used: a bulky first-generation model, and a slimline second-generation model, from different manufacturers. All cables were short (less than 1 metre), branded and marked as TB4 compliant. Three different SSDs were used:
- 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 an Orico case.
Benchmarking software used included my own Stibium, and AmorphousDiskMark. Stibium was used in two different test configurations:
- standard write-read test, consisting of writing 160 test files from 2 MB to 2 GB, totalling 53 GB, in randomised order, then reading those test files back in randomised order;
- random read or write test, consisting of 50 or 100 test files of random sizes in a uniform distribution from 2 MB to 2 GB.
When connected direct to the Mac, the SSDs returned the following results from the standard test:
- Disk A: read 2.7, write 2.2 GB/s
- Disk B: read 2.9, write 2.1 GB/s
- Disk C: read 2.7, write 2.8 GB/s.
When connected via either TB4 hub, all three SSDs retained similar read speeds to those measured when connected direct:
- Disk A: read 2.5-2.7 GB/s
- Disk B: read 2.6-2.7 GB/s
- Disk C: read 2.7-2.8 GB/s.
However, write speeds for Disks A and B were reduced to 1.5 GB/s consistently in every one of many measurements. Disk C retained a write speed of about 2.8 GB/s despite being connected through the same hub, often at the same time as either of the other two disks. This throttling in write speed is best seen in the measurements made during two random write tests on Disk A with Stibium, plotted together on the same chart.
The upper circles show 100 individual measurements of write speed on Disk A when connected direct to the Mac Studio, with the size of the file on the X axis, and the write speed on the Y axis. On both axes, 10^9 bytes is of course GB, so the consistent write speed found during this direct connection was 2.2 GB/s, as stated above.
The lower * symbols show 100 individual measurements of write speed on Disk A when connected through one of the hubs. Here, the consistent write speed is 1.5 GB/s, with just over 10% of results lower at around 1.2 GB/s. There are only three points from the upper series that lie within those from the lower series.
Where is the problem?
This throttling in write speed occurs independently of Mac: the same results are seen on a Mac Studio M1 Max and a MacBook Pro M1 Pro. It is independent of hub: the same results are seen on a first-generation TB4 hub, and one of the new second-generation TB4 models from a different manufacturer, and have been reported by others on at least two different hubs. It is independent of the test method: it’s seen in both types of test using Stibium, and when using AmorphousDiskMark, which uses completely different test methods. It is specific to the SSD: it doesn’t occur with Disk C, only with Disks A and B, and is entirely consistent.
This is therefore most probably a result of the interaction between the TB chips in the SSDs and hubs. Given that TB chips are designed and made by Intel, that doesn’t appear impressive.
- High-performance Thunderbolt 3 SSDs, with read speeds over 1.5 GB/s, connected through Thunderbolt 4 hubs, retain read speeds close to those when connected direct.
- Some high-performance TB3 SSDs show throttled write speeds when connected through TB4 hubs, resulting in write speeds of no more than 1.5 GB/s, equivalent to only two lanes of PCIe 3.0.
- Where write speed is important and use of a hub is intended, confirm that can be achieved by the combination of hub and SSD(s) intended for use.
- Those testing and reviewing TB4 hubs should assess their performance carefully.
- An explanation from Intel would seem appropriate.
I next turn to the complex issue of performance when reading or writing to more than one TB3 SSD at a time.