We’ve recently had an enormous amount of discussion about the performance of external SSDs. One issue that we’ve only touched upon is whether booting from a faster SSD results in a shorter time for an M1 Mac to start up.
I’ve just added to my stable of SSDs a combination that appears to reach maximum Thunderbolt 3 performance, as generally experienced. This is an ORICO SCM2T3 enclosure with a Samsung 980 PRO 2 TB NVMe SSD inside it, at a total cost of around $/€/£400, or $/€/£200 per TB. Orico doesn’t claim an ‘up to’ performance for its enclosure, but it is ‘certified by Intel’ and bears the Thunderbolt mark. The enclosure is claimed to have a JHL6340 (Intel Alpine Ridge) chipset supporting 4 PCIe Gen 3 lanes, which is reported as connecting at a speed of “up to 40Gb/s x1” and a link width of 2. It was tested while connected by a CalDigit 0.8 m Thunderbolt 4 cable to a Thunderbolt 4 port on a Mac Studio Max running macOS 12.4.
I ran two sets of transfer speed tests using Stibium. The first consisted of the standard set of 160 files, totalling just over 50 GB, and returned a write speed of 2.79 GB/s, and read of 2.83 GB/s. In the second, I doubled the number and total size of test files to exceed 100 GB, in case that might exhaust the SLC write cache or induce thermal throttling. As write speed remained at 2.78 GB/s and read at 2.75 GB/s, there was no evidence that larger load had any significant adverse effect on its performance.
For those collecting Amorphous results, the following screenshots tell all, first in MB/s, then in IOPS.
Following those tests, I installed macOS 12.4 on the SSD and measured the time taken for a cold boot, from when I pressed the Power button to the appearance of the Login window, which I compare with those from other disks:
- Internal SSD – 15 seconds (SSD read/write about 7 GB/s)
- Samsung 980 + ORICO case – 20 s (read/write about 2.8 GB/s)
- Samsung X5 – 20 s (read/write about 2 GB/s)
- Samsung 980 + Sabrent case – 18 s (read/write 1.5 GB/s)
- Crucial SATA/USB – 38 s (read/write 300-400 MB/s)
- Toshiba HDD – 100 s (read/write 70-150 MB/s).
Although increasing the transfer speed from the SATA SSD at up to 400 MB/s to more than 1 GB/s halved the time taken to boot, there’s no consistent improvement beyond that. This suggests that time to boot an M1 Mac from an external SSD is around 5 seconds longer than from the internal SSD, and independent of SSD performance above 1.5 GB/s. This most probably reflects the lengthier process, in which early boot takes place from the internal SSD anyway, then transfers to the external SSD only after its LocalPolicy has been checked and applied.
It also suggests a metric of MB/s per $/€/£ per TB, compared in the table below.
The price given for the Apple internal SSD is the cost difference between 1 and 2 TB in a Mac Studio Max. Under Apple’s current pricing, costs fall to $/€/£300 per TB with larger sizes. Read speeds are used here, as the Samsung X5 fills its SLC write cache quickly, resulting in poor write performance during more sustained tests. Costs include any enclosure required. Speed for the HD is given as the average across its range of 60-140 MB/s.
In terms of performance per cost, Apple internal SSDs are best value, followed by my new Samsung 980 PRO with the ORICO enclosure. The slower NVMe SSDs followed, with the Samsung X5 being marred by its poor sustained write performance (which isn’t taken into account here). SATA/USB SSDs follow, and 2 TB hard disks are in a distant place of their own.
I should have guessed that I’d be asked for Amorphous random benchmarks in IOPS for the four fastest SSDs. Here they are, given as read/write:
- Internal SSD – RND4K QD64 155,757/49,978 QD1 12,348/8,277
- Samsung 980 + ORICO case – RND4K QD64 280,821/76,497 QD1 15,325/9,323
- Samsung X5 – RND4K QD64 267,795/53,191 QD1 14,487/9,187
- Samsung 980 + Sabrent case – RND4K QD64 280,795/63,268 QD1 15,997/9,549
According to those, the fastest to boot should have been the two Samsung 980 SSDs, followed by the Samsung X5, with the Apple internal SSD the slowest of the four. I don’t think I’ll bother using Amorphous random tests given in IOPS again. By the way, while it may not be obvious to those who haven’t read Stibium’s documentation, both its write and read tests are automatically performed in random rather than sequential order.