The fastest USB storage options for Raspberry Pi

For years, I've been maintaining benchmarks for microSD cards on the Raspberry Pi, but I only spent a little time testing external USB storage, due to historic limitations with the Pi's USB 2.0 bus.

But the Pi 4 cleared away the limitations with a full-speed USB 3.0 bus offering much better performance, so I've done a lot of testing with USB boot, and with all the USB SSDs I had at my disposal. You can see some of those results in this blog post and video on booting a Pi 4 via USB.

After posting my tests concerning UASP support in USB SATA adapters, I got an email from Rob Logan mentioning the performance of some other types of drives he had with him. And he even offered to ship a few drives to me for comparisons!

There's also a video that accompanies this blog post, for the more visually-inclined:

So I took Rob up on the offer, and he sent me an XPG NVMe drive in a TDBT enclosure and an Arcanite AK58 USB 3.1 flash drive, touted by some as 'one of the fastest USB flash drives available'. I added a couple other drives to the test, pictured below:

Flash drives and USB SSD and NVMe from Arcanite Corsair TDBT XPG and Kingston with Raspberry Pi 4

Clockwise, from the Inateck case in the top middle:

I ran benchmarks on every one of these drives, testing their sequential read and write performance with hdparm and dd (to test large file operations), as well as their random 4K read and write performance with iozone (to test more general usage patterns when booting or running apps on a Pi).

Benchmark results

So how did all these drives perform? I ran all the tests booting the Pi from the device that was being tested, and I also ran the same benchmarks on my fastest microSD card, a Samsung Evo Plus.

Sequential I/O performance

Sequential performance of different USB drives on Raspberry Pi

The sequential results show a huge gap between the SSDs and fast new USB flash drives and the cheaper older-generation flash drives and the microSD card.

From what I've found, it seems like most USB flash drives perform the same as a microSD card in a USB adapter:

Wait... its all microsd - always has been astronaut meme

The Arcanite does well here, but it does lag a little bit behind the SSDs and even the Corsair GTX.

Random I/O Performance

Random IO performance of different USB drives on Raspberry Pi

Random IO performance paints a more complex picture. The performance of the older flash drives remains abysmal, with even the microSD card trouncing them in random 4K write performance.

But the Arcanite also falls off quite a bit in comparison to the SSDs and the Corsair. And the XPG NMVe drive is at least twice as fast as every other option when it comes to random write performance.

So overall, it looks like a decent quality NVMe drive and USB enclosure is going to give the best overall performance. And the Corsair GTX is by far the fastest USB flash drive I've ever tested.

But there's one other test I wanted to do before closing the book on performance, and that's a 10 GB file copy over the network.

This test doesn't sustain the maximum sequential throughput for the drive, but it does take a long time and tests how well the different devices handle heat from constant write activity.

Large file copy over network

NOTE: This graph is a little misleading. I believe that the scale should be 'Mbps', but I need to go back to my data and verify what exactly was measured, and re-generate the graph below. The relative scale is correct, however:

Large file network copy performance of different USB drives on Raspberry Pi

This shows some interesting results. The Arcanite and the SanDisk Ultra Fit perform much worse for long-duration file copies than all the rest.

The Arcanite was only a tiny bit slower than the SSDs and the Corsair in the quick sequential tests, and the Ultra Fit was actually slightly faster than the Ultra Flair.

Why do they perform so much worse in this benchmark?

Thermals

Well, I pulled out my Seek thermal camera and took a reading on the Ultra Fit:

Ultra fit gets very hot in thermal image

I put some thermal tape on the tiny bit of metal that was exposed when it's plugged in, and the temperature measured over 60°C (140°F)!

The Arcanite's plastic body didn't measure quite so hot, but that's just the problem—plastic is a good thermal insulator, and that's why you often see it used in coolers. Metal, on the other hand, is good at dispersing heat, but you have to have enough area for the metal to disperse the heat, or the drive is going to get really hot.

The Arcanite's plastic body traps the heat inside, which leads to overheating, while the Ultra Fit's tiny profile doesn't leave enough room to dissipate heat.

Compare that to the much beefier Corsair GTX, which is larger (lots more surface area) and made of solid metal. Even under heavy write load, the Corsair kept its cool at 36°C (<100°F):

Corsair Flash Voyager GTX thermal image

When you look at the benchmarks on the packaging, or even benchmarks posted to Amazon reviews from CrystalDiskMark, remember most of them don't reflect the true performance of these drives tested under real-world conditions.

SSDs and NVMe drives typically have a lot more surface area for heat dispersion, so they tend to perform more consistently since they can avoid overheating issues.

Most USB flash drives are designed more for compactness and convenience, and performance with general computing tasks or over long periods of time is usually more of an afterthought.

Usually, that is, unless we're talking about the Corsair—it's the first USB flash drive I've tested that compares favorably to USB SSDs!

Price Comparison

But what about price?

What drive gives the most value? After all, the Corsair is $55, while the cheapest option, the SanDisk USB Fit, is about $7. When you go to buy a USB drive for your Pi, you want the best overall value, and you might want to sacrifice a little performance for a lot in savings!

So comparing all these drives (using the price for the 128 GB version), I came up with these two graphs:

Price vs performance sequential IO for flash drives for Raspberry Pi

This graph compares how many dollars you have to spend per MB/sec on a large file copy. The Arcanite gives the best bang for your buck (assuming you're not constantly writing to it all day), with double the value of the SSDs or the Corsair GTX. The Kingston SSD comes in second... while the microSD and older SanDisk flash drives are a pretty poor choice when it comes to value for sequential access.

Price vs performance random IO for flash drives for Raspberry Pi

In this graph, showing value for random IO activity, the microSD card fares much better, but the older SanDisk flash drives are still a terrible value.

But the XPG NVMe (with an enclosure) becomes the best value, with the Kingston SSD and Corsair GTX in close pursuit. The Arcanite is a bit of a laggard, but it's still respectable with performance similar to the microSD drive.

Conclusion

What do these values mean? Should you get the XPG drive or the Corsair for the best raw performance? Or should you stick with the Arcanite, which gives the best bang for the buck for sequential performance (in some conditions)?

Well, that's impossible for me to answer. If you're going to store large files on the drive, and use it as a media server or NAS, then an Arcanite might be the best option. If you want to run applications or use the Pi as a desktop, the NVMe and an enclosure is probably the best option.

Or, if you want the most portable Pi possible, using the least amount of space and energy, you might be willing to sacrifice a little more performance and stick with a reliable microSD card.

Or if you need even more performance, you might want to look at a different single board computer that offers built-in SATA or NVMe support.

In the end, it's really up to how you want to use your Raspberry Pi, I just hope this helps make your decision a little easier.

More Information

I put all the raw performance data and benchmarks used in this issue on the Raspberry Pi Dramble issue tracker.

Also, Rob sent me many more results for other drives he tested, and was generous enough to allow me to share some of those results. They are summarized in the table below:

Drive hdparm - sequential 4k rand read 4k rand write
Inland Premium 256GB NVMe 3.0 x4 SSD 308.46 19.41 31.80
SanDisk 128GB Extreme Pro USB 3.1 Flash Drive 221.08 10.96 11.51
eMMC 5.1 module in USB 3.1 adapter 113.11 8.28 4.84
Netac Z8 250GB SM2258XT 299.05 20.48 31.47
Inland Pro 120GB PS3111-S11 313.05 18.59 24.66
eVtran 128GB ASM1153E 334.49 19.14 29.40

And he sent over a few pictures of some of the drives he tested, one of which I'll share here, for your reference:

Drives Rob Logan tested - actually only a few of the ones he tested

Comments

Hi Jeff, great analysis as ever - thank you. For some time now I have been intrigued by the possibility of creating the best value computation platform by booting from a very cheap source of memory (microSD) and then running an application or even the OS from a RAM disk on one of the more RAM-equipped Pi 4 models. I have seen two approaches to it so far. Firstly there's the more involved image generation using something like Alistair Buxton's rpi-ramdisk project (https://github.com/ali1234/rpi-ramdisk). The second, and perhaps more seductive, approach would be more like the one taken by HBFS Robotics where they run Chromium from a RAM disk (https://hbfsrobotics.com/blog/accelerating-raspberry-pi-4-applications-…). The trouble I have is that I can't see what would be involved in running another application (say, Octave) from a RAM disk because it doesn't have the kinds of files or installation that HBFS Robotics take advantage of in their example. I think this could be a real step forward in the value for money proposition of distributed computing but I'd love to know what you think.

Do you expect that the performance of the USB drives would improve if there was some active cooling around the 'stick'? In particular the Ultra Fit that reached some 60 degrees, might benefit from some indirect fan airflow.

Performance probably wouldn't improve, but consistency would (e.g. the Fit would work at it's highest speed for a huge file copy and not slow down a lot after a few minutes).

No, the Ultra Fit and Ultra Flair I have, do not get especially hot on the exterior, yet still slow down to a pathetic <15MB/s on large sequential writes during extended writes where they heat up internally, and in some cases even hiccup and freeze for a bit before continuing.

Some other people have observed their exterior getting hotter than mine from same generation so I can only assume the drive have some sort of linear regulator and the heat difference depends on the voltage of the USB port it is used on, despite the theory that they should all be dead on 5.0V, which we know is not always the case in reality.

At one point I considered modifying a hub to put a diode in series on the USB supply, to see if it would tolerate the lower voltage and if running cooler made it keep higher sustained read speeds, but since these are budget priced and nothing is going to make them perform like an Arcanite, or Sandisk Extreme Pro, I never bothered testing it.

Very interesting to see how different USB Drives compare. I don't understand the data for the Network transfer test. 232 MB/s is a lot to transfer over the Raspberry Pi's Gigabit Ethernet port. Are you sure it's MB and not Megabit? If so the numbers make sense.

With my own WD Green Internal SSD, Green, 240GB I measured 174 Mbit/s. No heating issue. Just Luke warm in a ORICO NGFF case.

Oh, yikes! I got that benchmark graphic all kinds of messed up. I'm not even sure what I did, but I had measured the time in seconds and the average rate in MB/s, so somewhere I must've screwed up the two or accidentally added a conversion to Mbps, as you say.

I will try to figure out what happened and get that benchmark updated.

Agreed some of those are too high to be mega*bytes* over GigE. But some seem too low for it to be mega*bits*. They look a little like expected mega*bytes* per second, but without network.

Hi Jeff,
you should give Samsungs T5 or T7 a try, hdparm tests giving 400+ MB read performance, I discovered those last year and have several, appears to be NVMe in a box, crazy fast storage. HTH

I just read a few Amazon reviews of the Corsair stick. They say the write performance comes from a 1.5 GB write cache, and data compression. Once the cache is exhausted, writing data becomes much slower. Especially when the data cannot be compressed much.

Thanks for the extensive comparison! I'm only getting started with SBCs and found this very helpful.