- Joined
- Feb 11, 2024
- Messages
- 116 (0.28/day)
- Location
- Canucksland
| System Name | Main / HTPC / Server |
|---|---|
| Processor | i5 14600K / Ryzen 5 2400G / i7 7700K |
| Motherboard | Z790 PRO RS / B450M Mortar / Z270 IX Code |
| Cooling | AS500 PLUS WH / Wraith Stealth / NH-D15 |
| Memory | 32GB 6000 C30 / 8GB 2666 C16 / 32GB 3000 C15 |
| Video Card(s) | GTX 780 (temporary) / iGPU / iGPU |
| Storage | (OS) 1600X 118GB / V200 120GB / 850 EVO 250GB |
| Display(s) | Predator XB271HU / KDL-55W950B / VH238H |
| Case | Eclipse P400S / LC13-BU / Define R5 |
| Audio Device(s) | Xonar U7 (HD 598) / Xfi Titanium (Azur 851A) |
| Power Supply | Prime Titanium 750W / M12II EVO 620W / AXi 860W |
In this post I’ll discuss my Rule of thumb, which to me is a quite interesting point that applies well to market advertising vs real world use, but also percolates to the internal SSD market to some extent. I’ll then make a Compilation of three USB 3.2 Gen 2 drives tested with CDM charts and assessment thereof regarding the subject at hand, and resume results in real use complete with some pics proving the highlighted aspects. Some Final considerations will follow.
THE RULE OF THUMB – DOWN BACKWARD PORTS AND/OR LESSER HARDWARE
To alleviate things a bit I’ll use USB 3.0 (5Gbps), 3.1 (10Gbps), and 3.2 (20Gbps) instead of 3.2 Gen “X” etc.
My initial query was in fact pretty simple: could I expect any 3.1 device to about “max out” the bandwidth of a 3.0 interface? Of course, not entirely max it out but behave like or even better than the fastest 3.0 devices out there – and they’re pretty rare. Not much market pushing them when the newer gens are out there with prices difficult to beat, so most cater to being dirt cheap instead, making them no better, and often worse, than my old Sandisk ones.
Sandisk Extreme 64GB (around 2013 IIRC...) – USB 3.0 interface – CDM results – No Fill
On my PC: i5 14600KF (6P+8E/20T mild OC @ 5.4/4.1 – Ring locked @ 4.7GHz) – 32GB DDR5 6000MHz...
On my laptop: i3 N305 (8C/8T @ 3.8GHz) – 8GB LPDDR5 4800MHz...
However, where I got a lot of relative answers here on TPU (“they’re plenty fast” and stuff like that), I did not get anything like a more precise figure about it. I wanted more precision: I do a lot of IT work across a lot of older hardware that will only have USB 3.0 ports at best while not being in the league of the hardware I use for myself. So... backwards compatibility/lesser hardware.
Otherwise, for exactly the same reason explained above I didn’t feel the need to pay extra for the 3.2 Gen 2x2 stuff or better; I absolutely needed a USB-A interface and wanted to avoid adapters if possible (easy to misplace); I ideally wanted a thumbdrive form factor as well because it’s less of a hassle, but I was amenable to go the external SSD route (and cables) otherwise. Capacity was not such of a big deal: I recycled an old 500GB Sata SSD and a loose Sata to USB 3.0 board from a now defunct external HDD, and still have a functional 2TB external HDD as well, and they’ve carried me okay so far, but force was to admit in 2025 that anything faster with 500GB-1TB capacity was not a luxury anymore – and would save me the hassle of the DIY SSD/USB 3.0 board thing.
Fortunately enough, the relative answers to my query got me into lateral thinking: my PC obviously DID offer a few USB 2.0 ports to test my Sandisk drives from... therefore, I could “relativize” too.
Sandisk Extreme 64GB – USB 2.0 interface (PC only) – CDM results – No Fill
The test worked even better than I expected: it sort of illuminated the whole actual market. My Sandisk 64GB, which does not come close to “maxing out” its native interface, got around 43MB/s sequential Reads at best on USB 2.0. So I started calculating... The theoretical bandwidth for USB 2.0 is 60MB/s. 43MB/s in turn is about 72% of that bandwidth for the Sandisk. What was that telling me?
Thinking back on the reviews I’ve read about USB 3.1 devices: theoretical bandwidth for them is 1250MB/s, but what you’ll mostly see of results however in synth benchmarks is an average 900MB/s.. A round 70% of their theoretical bandwidth gives 875MB/s. Same with USB 3.2 devices: theoretical bandwidth for them is 2500MB/s, average 1800MB/s – 70% being 1750MB/s indeed.
Therefore an interesting RULE OF THUMB: 70% is the conservative figure 3.1/3.2 devices can do from their native interface’s bandwidth within a realistic perspective. Where CDM is concerned , the average hover around 75% and the best of them around 85% in Peak tests. As it stands, most 3.1/3.2 devices use exactly 84% of their respective bandwidth as the basis of their advertising: up to 1050MB/s for 3.1 and up to 2100MB/s for 3.2. So 84% is the best figure manufacturers can expect of the interface (for Reads since Writes are commonly advertised lower) even across a relatively wide range of components and configurations. CDM Real World tests tend to go down around the 70% rule of thumb precisely. In real use scenario, even under ideal circumstances, drives seldom meet even CDM Real World speeds, but they do get very close to it, so the rule is mostly effective.
As an aside, if you look at nVME 4.0 SSDs for example, 70% of the theoretical bandwidth is 5600MB/s and 85% is 6800MB/s, both rather typical landmarks that circle the middle segment of that market pretty well. Of course, you can see up to +/- 90% of the bandwidth (7200MB/s) advertised there as they they benefit of a more direct interface with lesser latencies. Still in Real World CDM many of them tend to take a bigger hit vs. Peak results, which makes the 70% rule quite telling of what they can achieve in real use at best. Same kind of breakdown would apply to any internal SSD down to the good old Sata devices.
Further calculations: to the 43MB/s Sequential Reads I got with the Sandisk over USB 2.0 I got around 35MB/s Sequential Writes. And while that isn’t 70% of the bandwidth, it IS 80% of the Sequential Reads over USB 2.0, whereas from the native interface Sequential Writes were about 70% of Sequential Reads. RND4K stuff ranged from 80% to 99% of the native speeds over USB 2.0 and since the Sequential Reads are drastically lower there, it means that there is an overall tightening effect happening over Writes and RND4K speeds in general, and one that might lend over to lesser hardware as well (laptop’s results).
However, the Sandisk RND4K speeds are very low to begin with, so I would expect greater differences from faster devices when running from a backwards interface when the bandwidth is still largely sufficient to accomodate them, or from less powerful hardware particularly since I observed such effects first hand with internal SSDs/nVME drives as well – especially for RND4K Q32T1. But I still think the obvious tightening effect observed here should prevail with a backwards interface, leaving mostly the impact of lesser hardware regarding faster devices impossible to extrapolate from the Sandisk’s results alone as it was quite negligible there.
All that being said, and while I was willing to extrapolate that most devices would follow my rule of thumb from there, it’s still just benchmarks and theorectical numbers I’d been playing with so far. Real use stuff though can bring unforeseeable real use limits, as we’re about to see with the drives I tested... the kind of limits reviews don’t tend to test for, or when they do, don’t tend to push the test far enough to unveil them fully. But I did...
COMPILATION
All charts are No Fill and present each drive on the native interface and one gen backwards = USB 3.0. Same on my laptop except for the Kingston drive - I had legitimate reasons to shortcut testing it.
Real transfers on my PC only, all from the native interface unless specified otherwise (only one test with only the Patriot drive), and all Writes tests unless specified otherwise (one Reads test only for each the Patriot and SK Hynix). Even more tests were carried especially with the Patriot drive, but I’m resuming the essential here as Writes is where the problematic drives become obvious.
The source drives for all real transfers are two identical Silicon Power US75 2TB nVME SSDs which, while rather poor 4.0 drives in themselves, are still exceedingly faster than the USB drives tested and therefore did not affect their results. CDM for one of them:
A reminder of the main specs for both PC and laptop:
PC: i5 14600KF (6P+8E/20T mild OC @ 5.4/4.1 – Ring locked @ 4.7GHz) – 32GB DDR5 6000MHz
Laptop: i3 N305 (8C/8T @ 3.8GHz) – 8GB LPDDR5 4800MHz
The laptop obviously constitutes the “lesser hardware” scenario used in the CDM assessments. It’s somewhat better than many different/older configurations I encounter, but not SO MUCH better, and not ALL of them. Its specs give a correct idea of up to 10 years old hardware in average, but not of the direst extremities of such a time span... nor I believe is it essential to the point since it’d still be eons better than being stuck with truly ancient hardware over USB 2.0.
PATRIOT SUPERSONIC RAGE PRIME 500GB – ADVERTISED AT 600MB/s READS (WRITES UNDISCLOSED)
Bought for its cheap price and retractable thumbdrive form. Not the speediest 3.1 drive, but the advertised speed suited me okay.
CDM CHART
CDM ASSESSMENT – BACKWARDS INTERFACE/LESSER HARDWARE
The Patriot fully meets its claim from its native 3.1 interface looking at Peak Sequentials in CDM, and it does pretty good of the Real World tests there as well: more than 80% of Sequential Reads compared to Peak results for the very same Sequential Writes results, while at a glance all RND4K results are par for the course with most USB 3.1 devices. The drive also fully meets my rule of thumb about backwards compatibility, where we also find altogether tight RND4K results as well: Peak Sequentials are 75% of the bandwidth overall; Real World Sequentials follow the same exact trend than over USB 3.1 from there; Peak RND4K about 75% overall and Real World RND4K 94-99% of USB 3.1.
It was expected that, notwithstanding Sequentials that are firstly impacted by bandwidth limitation (though in this case barely), RND4K Q32T1 would be where the slower interface speeds would carry most impact even if in their case the theoretical bandwidth “normalized” by rule of thumb is largely sufficient.
The impact of lesser hardware is obvious in parts of the results: Real World Sequentials overall, Peak RND4K overall and Real World RND4K Writes on my laptop come down to about 90%, 75% and 50% respectively of the results on my PC both from USB 3.1 and 3.0; however Peak Sequentials overall and Real World RND4K Reads are within a 5% margin at worst. RND4K are not typcally a big concern for external drives, so for typical use indeed I’d say the impact of lesser hardware with this drive is largely negligible.
The tightening effect observed with the Sandisk Extreme going down one gen holds true for the Patriot and mostly carries over lesser hardware (relative impact is no worse) where I feel it’s very important since lesser hardware often equals USB 3.0 ports also.
REAL USE SCENARIO
Writes: No Fill...
...for a 421GB movie transfer: drop from around 550MB/s to around 220MB/s at 25% = some 105GB filled and took another hit to around 20MB/s at 28% = some 118GB filled. Hovered around 20MB/s for the rest of the transfer.
Please note that 28% of that transfer, which is around 118GB as stated, corresponds to about 25% of the capacity filled.
USB 3.0 – Writes: No Fill...
...for the same 421GB transfer: drop from around 360MB/s to around 230MB/s at 25% and took another hit to around 20MB/s at 28%. Hovered there for the rest = exactly same behavior to same capacity filled than on USB 3.1.
Other tests were carried on USB 3.1 to prove the above, or filling my old Sandisk Extreme (USB 3.0) to see if a similar behavior could be observed (the Sandisk held steady throughout at +/- 175MB/s) etc. Which leads to the more interesting tests below, all of them back on USB 3.1.
Writes: Filled to a crawl / cool down / second transfer...
... for an additional 115GB as to verify if it recovered in further uses within a data set that it could do well, therefore if it MIGHT avoid the crawl altogether when only adding data in smaller batches than its cache size (regular everyday use). This is the... hypothesis most reviews that do a sustained Writes test at all, and therefore hit the cache wall, tend to relay in assessing the drive.
The add-on started around full speed for a couple seconds, then went down to the 20MB/s crawl within 5% of the transfer done. Therefore the hypothesis that the hit sustained by most drives in a sustained Writes test is “temporary” or conditional to bigger transfers only. In fact, we’ll see similar behavior with the next drive, and I’ll make a case in the Final word that I do not think the cache size is the sole factor at play there.
Reads: Filled up to 421GB...
...copying them movies back into a “Test Folder” on the other SP US75: drop from around 420MB/s to around 360MB/s at 25% (some 105GB read back), then to around 290MB/s at 68% (some 285GB read back) and hovering around 280-300MB/s until the end. So in sustained Reads, while we have nothing as dramatic as in Writes, the first hit at 25% precisely of the very same amount of the very same files was quite telling, and the 290MB/s average sustained for the last third of the test is disappointing. Of course it’s still quite usable there, but it circles back to the Writes problem: if you need to copy as much from it, you’d need to fill it as much first.
Writes: Varied file types with LOTS of 1KB files on top of various stuff from 50KB to 4GB/No Fill
...for a 280GB transfer: unrecoverable drop to the 20MB/s crawl verified around 47% of the transfer = 132GB filled. Let it run until 53% to make sure it would not climb back up. So a bit more leeway with varied file sizes before the crawl, but nothing to make me reconsider. Also, the next – and last – test from there proved me right, see below.
Writes: Adding movies on top of aborted varied test/Filled up to 150GB...
... for a small 15GB movie transfer: went to a crawl well within 10% done. Let it run some more to make sure, and at 39% it dropped to an unprecedented 8MB/s crawl. Stopped it right there. No further testing.
Bargained successfully for a full in-store credit without sending it back. Partitioned it like a 128GB drive is under Windows for long term use. Glad I managed that bargain too: as a free “128GB” thumbdrive it’s excellent... the keyword being “free”.
KINGSTON XS1000R 1000GB – ADVERTISED AT 1050MB/s READS & 950MB/s WRITES – NOT TESTED ON LAPTOP
Bought mostly because it was amongst the least expensive 1TB exernal SSD with a “good value” reviewing from TPU.
CDM CHART
CDM ASSESSMENT – BACKWARDS INTERFACE/LESSER HARDWARE
The Kingston fully meets its claim from its native 3.1 interface looking at Peak Sequentials in CDM, and it does pretty good of the Real World tests there as well: around 90% of Sequentials overall compared to Peak results, while at a glance all RND4K results are par for the course with most USB 3.1 devices. The drive also fully meets my rule of thumb about backwards compatibility, where we also find altogether tight RND4K results as well: Peak Sequentials are 75% of the bandwidth overall; Real World Sequentials are 95% overall from there; Peak RND4K about 80% overall and Real World RND4K 92-94% of USB 3.1.
It was expected that, notwithstanding Sequentials that are firstly impacted by bandwidth limitation, RND4K Q32T1 would be where the slower interface speeds would carry most impact even if in their case the theoretical bandwidth “normalized” by rule of thumb is largely sufficient.Those results similar enough to the Patriot drive that, even without testing the Kingston on my laptop, I suspect we’d roughly see the same kind of hit from lesser hardware, which would bring the same conclusion: its impact largely negligible.
The tightening effect observed with the Sandisk Extreme AND Patriot going down one gen holds true for the Kingston drive as well, and should fully carry over lesser hardware (relative impact no worse) where I feel it’s most important.
REAL USE SCENARIO
*Please note that, while still video files, this data set is entirely different from the one used with the Patriot drive.
Writes: No Fill...
...for a 841GB movie transfer: drop from around 740MB/s to around 150MB/s at 22% = some 185GB filled and took another hit to around 100MB/s at 28% = some 235GB filled. Hovered around 100MB/s for the rest of the transfer.
Almost exactly mirroring the Patriot’s two-stages unfolding into... well not quite a crawl... yet! Especially, note that 28% of this transfer, which is around 235GB filled, again corresponds to about 25% of its capacity filled.
Writes: Filled up to 841GB
...for a 80GB transfer filling it up to the brim: drop from around 700MB/s to around 30MB/s at 35% = an additional 28GB = some 869GB filled (93%). Did not recover. Worse drop than in sustained Writes and very similar there to the Patriot’s crawl.
Writes: Filled to 421GB (reciprocating amount of data used for Patriot drive as tested as well as below 50% capacity filled)...
...for an additional 182GB movie transfer: drop from around 700MB/s to 30MB/s average after 13% = an additional 24GB = some 445GB filled. Did not recover. Note that at this point we’re still below 50% of the capacity filled, so it doesn’t have anything to do with an effect of the drive being close to being filled up as it MIGHT have been in the last test.
Between this test and the one before, it seems that when filled to more than 25% of its capacity, the Kingston drive has around 25-30GB of data it can add near top speed before coming to a crawl that belies even the +/- 100MB/s it could keep up in a sustained Writes test, and while I cannot explain the latter, the 30MB/s itself in subsequent tests looks just like the Patriot’s crawl and general unability to recover, which I considered quite telling. Of course as you can see, I do not have the same burden of proof than with the Patriot drive. I was just looking out for the signs, and I had them already.
Instead I went to format it thinking I’d do an add-on test with the drive filled up until the full drop down 100MB/s, as it seemed to have the same breaking point (from the first test) than the Patriot drive, which I pegged to be around 25% of the capacity filled. Just to see how it would behave in subsequent uses after a cool down and with just a 100GB add-on...
First time formatting through regular OS option...
...bricked/went into RAW partition. Targeted Web search on this series reveals a tendency to various such failures reported by many users. No further testing – no curiosity about its Reads potential, and quite out of patience with such inconsistencies. Sent back as is for a full refund this time – had no difficulty at all obtaining it. Restablishing it from the RAW state would have been easy but was NOT compelling just to prove a point about a drive I had no intentions to keep anyway. Take a look at the FINAL WORD.
SK HYNIX X31 1024GB – ADVERTISED AT 1050MB/s READS & 1000MB/s WRITES
Bought because Amazon had it for the cheapest ever money for a 1TB drive, even vs. the cheapest “no-name” China brands – and obviously, it was one of the 3.1 drives with some of the best reviews and (whenever tested) sustained Writes speed out there.
CDM CHART
CDM ASSESSMENT – BACKWARDS INTERFACE/LESSER HARDWARE
The SK Hynix fully meets its claim from its native 3.1 interface looking at Peak Sequentials in CDM, and even exceeds them for Writes which keep real close to Reads speed. It does pretty good of the Real World tests there as well: 85% of Peak Sequentials overall. However, where at a glance most RND4K Reads results are par for the course with most USB 3.1 devices, there are some quirks with RND4K Real World Writes where they were faster on USB 3.0 than USB 3.1 on my PC, yet either way 25-45% slower than the other drives tested on my PC . However, it’s Read speeds there were consistently 10-20% faster than the other drives, especially on my laptop. Peak RND4K were overall relatively stronger as well, and again especially on my laptop. All in all, I prefer its strenght and consistency over lesser hardware – and one gen down – to whatever quirk/lesser RND4K Real World Writes ahead.
The drive also fully meets my rule of thumb about backwards compatibility: Peak Sequentials are 75% of the bandwidth overall; Real World Sequentials Reads stronger than the other drives for similar Writes, to 90% overall of the Peak Sequentials; Peak RND4K about 77.5% overall and Real World RND4K 94-99(130)% of USB 3.1. The SK Hynix does the best of it for all drives tested, counting with or without the quirk between PC and laptop. It was expected that, notwithstanding Sequentials that are firstly impacted by bandwidth limitation, RND4K Q32T1 would be where the slower interface speeds would carry most impact even if in their case the theoretical bandwidth “normalized” by rule of thumb is largely sufficient.
The impact of lesser hardware is more than ever nothing to worry about with the X31: Peak RND4K takes the worst hit with values ranging from 70-90% (which is still excellent) of the results from my PC, and everything else is 90-100(130)% with a heavy concentration at the 95%+ mark. For typical use the impact is really negligible here where the SK Hynix really shines through.
The tightening effect observed with all other drives going down one gen obviously holds true here but much more importantly carries quite spectacularly over lesser hardware, making the X31 a highly consistent and seamless as possible performer in all uses.
REAL USE SCENARIO
*Please note that this data set is the same exact one used with the Kingston drive.
Writes: No Fill...
...for a 841GB movie transfer: drop from around 750MB/s to around 200MB/s at 63% = some 530GB filled = more than even 2TB drives with pSLC cache would sustain at full speed. Then the X31 recovered to full speed around 76% = some 640GB filled. Kept around 750MB/s until the transfer was nearly done = 510MB/s average.
Writes: Filled up to 841GB
...for an additional 111GB movie transfer: around 750MB/s without any drop. So THIS drive DOES recover, even further filling it within 0.5% of its full capacity with a rather additional large data set.
From a curiosity standpoint, and while I did no sustained Reads test, a quick cut/paste test back to the source drive with some of the movies showed nearly 900MB/s speed. Even if it were to drop a bit after some 25% of its capacity like the Patriot drive did, and up to 530GB as per the SK Hynix behavior in sustained Writes, I had no doubt it’d still be satisfying. I rather thought that the Reads test we’ll see below would be more interesting to illustrate the capacities of this so far very promising drive.
Writes: Varied file types with LOTS of 1KB files on top of various stuff from 50KB to 4GB/No Fill
... for a 280GB transfer: about 13 minutes = 370MB/s average and no permanent drop if a lot of expected dips. To keep things fair and square, the Patriot could do about the same speed in this test – until it went into an unrecoverable 20MB/s crawl before half was done.
Reads: Varied file types with LOTS of 1KB files on top of various stuff from 50KB to 4GB/Filled 280GB
...copying them files back into a “Test Folder” on the source drive they came from. Some 600MB/s average with just the expected dips from the varied file sizes, but no drop. Of course that’s something the Patriot drive wouldn’t be able to pull even before it would inevitably slow down like it did in a fully Sequential Reads test.
No further testing. A clear winner and keeper that comes with a bit of extra space over most “1TB” drives and keeps full speed for sustained Writes longer than most “2TB” drives with pSLC cache can while comparing favorably in cost/capacity buying two of them – and recovering from a speed drop while in sustained use! Also brings the possibility of similar advantages in a thumbdrive (admittedly rather bulky) form with the T31.
FINAL CONSIDERATIONS: THE RULE OF THUMB & CACHE/FILL LIMITATIONS
My initial findings and extrapolating with my old Sandisk Extreme drive from USB 3.0 to USB 2.0 proved right with all three USB 3.1 drives tested – where CDM is concerned and from a USB 3.1 to USB 3.0 port in their case. The tightening effect proved right as well when going backwards one gen. Lesser hardware would not be a particularly noticeable drawback to any of the three drives tested in typical use, but especially the SK Hynix was a consistent drive in all tests that really shined when benchmarked even from a 3.0 port on my laptop – or in real use.
While the sample tested here is not representative of the whole market, two drives from different brands and form factor and components behaved relatively similarly in that, once filled to around 25% of their capacity they are permenantly crippled until emptied out below that point or formatted anew altogether. The Kingston was a bit more forthcoming than the Patriot, in that it could do some 25GB add-ons at best speeds from there (while the Patriot just about instantly went to a crawl), but then went way below what it could even hold in a sustained Writes test, almost exactly mirroring the Patriot’s crawling speed.
With the Patriot was also found that Reads would take a noticeable hit whenever reading back more than 25% of its capacity as well, but to a lesser extent than with Writes, and reasonably fast even there. However the point stands that if you need to Read as much you need to have filled any of them as much in the first place, so it circles back to the Writes problem.
The only notable difference between either the Patriot or the Kingston, and the SK Hynix which behaved splendidly in a sustained Writes test AND could recover in subsequent uses when filled some more even up to 99%+ of its capacity is that the Hynix has a DRAM cache while the other two only rely on a pSLC cache. While I do not have enough drives tested to make it into a general rule, nor did I test the Kingston thoroughly enough to make sure, I can still issue a very severe warning about drives without a DRAM cache especially, but also that it seems as much of a thing with capacity filled up to 25%+ than strictly a cache thing. I’d test any such drive at least thoroughly enough to make sure no such behavior is readily observable.
Therefore I’m not entirely sure what causes the effects I saw/where cache plays into it. A good exemple of this is just comparing the Kingston XS1000 1TB I tested vs. what TPU said in a review of the 2TB unit of the same Kingston series:
“Sustained write speeds of the Kingston SSD aren't that impressive. With just 180 MB/s to fill the whole disk, it's slower than many of its competitors, but considerably faster than the Crucial X8. If you plan on copying hundreds of gigabytes per session without pause, then look elsewhere. If your working sets are smaller, up to 100 GB writes, or so, or you have pauses in transfers that give the drive room to breathe, the XS1000 will run much faster, reaching around 650 MB/s—the pseudo-SLC cache size is 120 GB, which is a reasonable size.” https://www.techpowerup.com/review/kingston-xs1000/
Obviously TPU reviewers know more than I do, so on the one hand IF the pSLC cache of the Kingston 2TB is 120GB, surely the 1TB I tested has at most an equal cache. Yet, the 2TB tested by TPU would hit that 180MB/s wall in a sustained Writes test just about when the 120GB cache was out of the equation, whereas mine was perfectly fine until 185GB filled, and only got to its slowest average speed of 100MB/s after 235GB filled.
Then again, TPU remarked that the drive would run much faster with smaller working sets “up to 100GB or so” or with pauses to “give the drive room to breath” between transfers, but mine could only sustain about 25GB add-ons at full speed before dropping even lower than 100MB/s – the so-called crawl I’ve been talking about, of which I never heard a word in many reviews, even those that do a sustained Writes test, and also relay the hypothesis that the drive would regain full speed with smaller data batches or with breathing room between use. Reminder: the Patriot didn’t have nearly as much breathing room whatsoever into that crawl.
Of course, there might be good reasons why TPU got what they got and why it’s so different from what I got with quick testing the Kingston XS1000, and I’d sure like to have @W1zzard POV on my findings, but from where I stand I’d be particularly wary of bigger drives because they could very possibly be even more useless as per capacity than the ones I tested here. But more generally speaking, and reiterating the point, TPU is not the only source that do a sustained Writes test and note a drop equivalent to the cache size filled, nor the only source to maintain that the drive would likely recover with breathing room or smaller data sets.
Hope I have contributed enough for folks to make up their own mind about it. Hopefully, some will try similar experimenting with their drives and prove me wrong, perhaps observe different symptoms, or confirm in some cases that such a problem exists. More data around this would surely be appreciable.
THE RULE OF THUMB – DOWN BACKWARD PORTS AND/OR LESSER HARDWARE
To alleviate things a bit I’ll use USB 3.0 (5Gbps), 3.1 (10Gbps), and 3.2 (20Gbps) instead of 3.2 Gen “X” etc.
My initial query was in fact pretty simple: could I expect any 3.1 device to about “max out” the bandwidth of a 3.0 interface? Of course, not entirely max it out but behave like or even better than the fastest 3.0 devices out there – and they’re pretty rare. Not much market pushing them when the newer gens are out there with prices difficult to beat, so most cater to being dirt cheap instead, making them no better, and often worse, than my old Sandisk ones.
Sandisk Extreme 64GB (around 2013 IIRC...) – USB 3.0 interface – CDM results – No Fill
On my PC: i5 14600KF (6P+8E/20T mild OC @ 5.4/4.1 – Ring locked @ 4.7GHz) – 32GB DDR5 6000MHz...
On my laptop: i3 N305 (8C/8T @ 3.8GHz) – 8GB LPDDR5 4800MHz...
However, where I got a lot of relative answers here on TPU (“they’re plenty fast” and stuff like that), I did not get anything like a more precise figure about it. I wanted more precision: I do a lot of IT work across a lot of older hardware that will only have USB 3.0 ports at best while not being in the league of the hardware I use for myself. So... backwards compatibility/lesser hardware.
Otherwise, for exactly the same reason explained above I didn’t feel the need to pay extra for the 3.2 Gen 2x2 stuff or better; I absolutely needed a USB-A interface and wanted to avoid adapters if possible (easy to misplace); I ideally wanted a thumbdrive form factor as well because it’s less of a hassle, but I was amenable to go the external SSD route (and cables) otherwise. Capacity was not such of a big deal: I recycled an old 500GB Sata SSD and a loose Sata to USB 3.0 board from a now defunct external HDD, and still have a functional 2TB external HDD as well, and they’ve carried me okay so far, but force was to admit in 2025 that anything faster with 500GB-1TB capacity was not a luxury anymore – and would save me the hassle of the DIY SSD/USB 3.0 board thing.
Fortunately enough, the relative answers to my query got me into lateral thinking: my PC obviously DID offer a few USB 2.0 ports to test my Sandisk drives from... therefore, I could “relativize” too.
Sandisk Extreme 64GB – USB 2.0 interface (PC only) – CDM results – No Fill
The test worked even better than I expected: it sort of illuminated the whole actual market. My Sandisk 64GB, which does not come close to “maxing out” its native interface, got around 43MB/s sequential Reads at best on USB 2.0. So I started calculating... The theoretical bandwidth for USB 2.0 is 60MB/s. 43MB/s in turn is about 72% of that bandwidth for the Sandisk. What was that telling me?
Thinking back on the reviews I’ve read about USB 3.1 devices: theoretical bandwidth for them is 1250MB/s, but what you’ll mostly see of results however in synth benchmarks is an average 900MB/s.. A round 70% of their theoretical bandwidth gives 875MB/s. Same with USB 3.2 devices: theoretical bandwidth for them is 2500MB/s, average 1800MB/s – 70% being 1750MB/s indeed.
Therefore an interesting RULE OF THUMB: 70% is the conservative figure 3.1/3.2 devices can do from their native interface’s bandwidth within a realistic perspective. Where CDM is concerned , the average hover around 75% and the best of them around 85% in Peak tests. As it stands, most 3.1/3.2 devices use exactly 84% of their respective bandwidth as the basis of their advertising: up to 1050MB/s for 3.1 and up to 2100MB/s for 3.2. So 84% is the best figure manufacturers can expect of the interface (for Reads since Writes are commonly advertised lower) even across a relatively wide range of components and configurations. CDM Real World tests tend to go down around the 70% rule of thumb precisely. In real use scenario, even under ideal circumstances, drives seldom meet even CDM Real World speeds, but they do get very close to it, so the rule is mostly effective.
As an aside, if you look at nVME 4.0 SSDs for example, 70% of the theoretical bandwidth is 5600MB/s and 85% is 6800MB/s, both rather typical landmarks that circle the middle segment of that market pretty well. Of course, you can see up to +/- 90% of the bandwidth (7200MB/s) advertised there as they they benefit of a more direct interface with lesser latencies. Still in Real World CDM many of them tend to take a bigger hit vs. Peak results, which makes the 70% rule quite telling of what they can achieve in real use at best. Same kind of breakdown would apply to any internal SSD down to the good old Sata devices.
Further calculations: to the 43MB/s Sequential Reads I got with the Sandisk over USB 2.0 I got around 35MB/s Sequential Writes. And while that isn’t 70% of the bandwidth, it IS 80% of the Sequential Reads over USB 2.0, whereas from the native interface Sequential Writes were about 70% of Sequential Reads. RND4K stuff ranged from 80% to 99% of the native speeds over USB 2.0 and since the Sequential Reads are drastically lower there, it means that there is an overall tightening effect happening over Writes and RND4K speeds in general, and one that might lend over to lesser hardware as well (laptop’s results).
However, the Sandisk RND4K speeds are very low to begin with, so I would expect greater differences from faster devices when running from a backwards interface when the bandwidth is still largely sufficient to accomodate them, or from less powerful hardware particularly since I observed such effects first hand with internal SSDs/nVME drives as well – especially for RND4K Q32T1. But I still think the obvious tightening effect observed here should prevail with a backwards interface, leaving mostly the impact of lesser hardware regarding faster devices impossible to extrapolate from the Sandisk’s results alone as it was quite negligible there.
All that being said, and while I was willing to extrapolate that most devices would follow my rule of thumb from there, it’s still just benchmarks and theorectical numbers I’d been playing with so far. Real use stuff though can bring unforeseeable real use limits, as we’re about to see with the drives I tested... the kind of limits reviews don’t tend to test for, or when they do, don’t tend to push the test far enough to unveil them fully. But I did...
COMPILATION
All charts are No Fill and present each drive on the native interface and one gen backwards = USB 3.0. Same on my laptop except for the Kingston drive - I had legitimate reasons to shortcut testing it.
Real transfers on my PC only, all from the native interface unless specified otherwise (only one test with only the Patriot drive), and all Writes tests unless specified otherwise (one Reads test only for each the Patriot and SK Hynix). Even more tests were carried especially with the Patriot drive, but I’m resuming the essential here as Writes is where the problematic drives become obvious.
The source drives for all real transfers are two identical Silicon Power US75 2TB nVME SSDs which, while rather poor 4.0 drives in themselves, are still exceedingly faster than the USB drives tested and therefore did not affect their results. CDM for one of them:
A reminder of the main specs for both PC and laptop:
PC: i5 14600KF (6P+8E/20T mild OC @ 5.4/4.1 – Ring locked @ 4.7GHz) – 32GB DDR5 6000MHz
Laptop: i3 N305 (8C/8T @ 3.8GHz) – 8GB LPDDR5 4800MHz
The laptop obviously constitutes the “lesser hardware” scenario used in the CDM assessments. It’s somewhat better than many different/older configurations I encounter, but not SO MUCH better, and not ALL of them. Its specs give a correct idea of up to 10 years old hardware in average, but not of the direst extremities of such a time span... nor I believe is it essential to the point since it’d still be eons better than being stuck with truly ancient hardware over USB 2.0.
PATRIOT SUPERSONIC RAGE PRIME 500GB – ADVERTISED AT 600MB/s READS (WRITES UNDISCLOSED)
Bought for its cheap price and retractable thumbdrive form. Not the speediest 3.1 drive, but the advertised speed suited me okay.
CDM CHART
CDM ASSESSMENT – BACKWARDS INTERFACE/LESSER HARDWARE
The Patriot fully meets its claim from its native 3.1 interface looking at Peak Sequentials in CDM, and it does pretty good of the Real World tests there as well: more than 80% of Sequential Reads compared to Peak results for the very same Sequential Writes results, while at a glance all RND4K results are par for the course with most USB 3.1 devices. The drive also fully meets my rule of thumb about backwards compatibility, where we also find altogether tight RND4K results as well: Peak Sequentials are 75% of the bandwidth overall; Real World Sequentials follow the same exact trend than over USB 3.1 from there; Peak RND4K about 75% overall and Real World RND4K 94-99% of USB 3.1.
It was expected that, notwithstanding Sequentials that are firstly impacted by bandwidth limitation (though in this case barely), RND4K Q32T1 would be where the slower interface speeds would carry most impact even if in their case the theoretical bandwidth “normalized” by rule of thumb is largely sufficient.
The impact of lesser hardware is obvious in parts of the results: Real World Sequentials overall, Peak RND4K overall and Real World RND4K Writes on my laptop come down to about 90%, 75% and 50% respectively of the results on my PC both from USB 3.1 and 3.0; however Peak Sequentials overall and Real World RND4K Reads are within a 5% margin at worst. RND4K are not typcally a big concern for external drives, so for typical use indeed I’d say the impact of lesser hardware with this drive is largely negligible.
The tightening effect observed with the Sandisk Extreme going down one gen holds true for the Patriot and mostly carries over lesser hardware (relative impact is no worse) where I feel it’s very important since lesser hardware often equals USB 3.0 ports also.
REAL USE SCENARIO
Writes: No Fill...
...for a 421GB movie transfer: drop from around 550MB/s to around 220MB/s at 25% = some 105GB filled and took another hit to around 20MB/s at 28% = some 118GB filled. Hovered around 20MB/s for the rest of the transfer.
Please note that 28% of that transfer, which is around 118GB as stated, corresponds to about 25% of the capacity filled.
USB 3.0 – Writes: No Fill...
...for the same 421GB transfer: drop from around 360MB/s to around 230MB/s at 25% and took another hit to around 20MB/s at 28%. Hovered there for the rest = exactly same behavior to same capacity filled than on USB 3.1.
Other tests were carried on USB 3.1 to prove the above, or filling my old Sandisk Extreme (USB 3.0) to see if a similar behavior could be observed (the Sandisk held steady throughout at +/- 175MB/s) etc. Which leads to the more interesting tests below, all of them back on USB 3.1.
Writes: Filled to a crawl / cool down / second transfer...
... for an additional 115GB as to verify if it recovered in further uses within a data set that it could do well, therefore if it MIGHT avoid the crawl altogether when only adding data in smaller batches than its cache size (regular everyday use). This is the... hypothesis most reviews that do a sustained Writes test at all, and therefore hit the cache wall, tend to relay in assessing the drive.
The add-on started around full speed for a couple seconds, then went down to the 20MB/s crawl within 5% of the transfer done. Therefore the hypothesis that the hit sustained by most drives in a sustained Writes test is “temporary” or conditional to bigger transfers only. In fact, we’ll see similar behavior with the next drive, and I’ll make a case in the Final word that I do not think the cache size is the sole factor at play there.
Reads: Filled up to 421GB...
...copying them movies back into a “Test Folder” on the other SP US75: drop from around 420MB/s to around 360MB/s at 25% (some 105GB read back), then to around 290MB/s at 68% (some 285GB read back) and hovering around 280-300MB/s until the end. So in sustained Reads, while we have nothing as dramatic as in Writes, the first hit at 25% precisely of the very same amount of the very same files was quite telling, and the 290MB/s average sustained for the last third of the test is disappointing. Of course it’s still quite usable there, but it circles back to the Writes problem: if you need to copy as much from it, you’d need to fill it as much first.
Writes: Varied file types with LOTS of 1KB files on top of various stuff from 50KB to 4GB/No Fill
...for a 280GB transfer: unrecoverable drop to the 20MB/s crawl verified around 47% of the transfer = 132GB filled. Let it run until 53% to make sure it would not climb back up. So a bit more leeway with varied file sizes before the crawl, but nothing to make me reconsider. Also, the next – and last – test from there proved me right, see below.
Writes: Adding movies on top of aborted varied test/Filled up to 150GB...
... for a small 15GB movie transfer: went to a crawl well within 10% done. Let it run some more to make sure, and at 39% it dropped to an unprecedented 8MB/s crawl. Stopped it right there. No further testing.
Bargained successfully for a full in-store credit without sending it back. Partitioned it like a 128GB drive is under Windows for long term use. Glad I managed that bargain too: as a free “128GB” thumbdrive it’s excellent... the keyword being “free”.
KINGSTON XS1000R 1000GB – ADVERTISED AT 1050MB/s READS & 950MB/s WRITES – NOT TESTED ON LAPTOP
Bought mostly because it was amongst the least expensive 1TB exernal SSD with a “good value” reviewing from TPU.
CDM CHART
CDM ASSESSMENT – BACKWARDS INTERFACE/LESSER HARDWARE
The Kingston fully meets its claim from its native 3.1 interface looking at Peak Sequentials in CDM, and it does pretty good of the Real World tests there as well: around 90% of Sequentials overall compared to Peak results, while at a glance all RND4K results are par for the course with most USB 3.1 devices. The drive also fully meets my rule of thumb about backwards compatibility, where we also find altogether tight RND4K results as well: Peak Sequentials are 75% of the bandwidth overall; Real World Sequentials are 95% overall from there; Peak RND4K about 80% overall and Real World RND4K 92-94% of USB 3.1.
It was expected that, notwithstanding Sequentials that are firstly impacted by bandwidth limitation, RND4K Q32T1 would be where the slower interface speeds would carry most impact even if in their case the theoretical bandwidth “normalized” by rule of thumb is largely sufficient.Those results similar enough to the Patriot drive that, even without testing the Kingston on my laptop, I suspect we’d roughly see the same kind of hit from lesser hardware, which would bring the same conclusion: its impact largely negligible.
The tightening effect observed with the Sandisk Extreme AND Patriot going down one gen holds true for the Kingston drive as well, and should fully carry over lesser hardware (relative impact no worse) where I feel it’s most important.
REAL USE SCENARIO
*Please note that, while still video files, this data set is entirely different from the one used with the Patriot drive.
Writes: No Fill...
...for a 841GB movie transfer: drop from around 740MB/s to around 150MB/s at 22% = some 185GB filled and took another hit to around 100MB/s at 28% = some 235GB filled. Hovered around 100MB/s for the rest of the transfer.
Almost exactly mirroring the Patriot’s two-stages unfolding into... well not quite a crawl... yet! Especially, note that 28% of this transfer, which is around 235GB filled, again corresponds to about 25% of its capacity filled.
Writes: Filled up to 841GB
...for a 80GB transfer filling it up to the brim: drop from around 700MB/s to around 30MB/s at 35% = an additional 28GB = some 869GB filled (93%). Did not recover. Worse drop than in sustained Writes and very similar there to the Patriot’s crawl.
Writes: Filled to 421GB (reciprocating amount of data used for Patriot drive as tested as well as below 50% capacity filled)...
...for an additional 182GB movie transfer: drop from around 700MB/s to 30MB/s average after 13% = an additional 24GB = some 445GB filled. Did not recover. Note that at this point we’re still below 50% of the capacity filled, so it doesn’t have anything to do with an effect of the drive being close to being filled up as it MIGHT have been in the last test.
Between this test and the one before, it seems that when filled to more than 25% of its capacity, the Kingston drive has around 25-30GB of data it can add near top speed before coming to a crawl that belies even the +/- 100MB/s it could keep up in a sustained Writes test, and while I cannot explain the latter, the 30MB/s itself in subsequent tests looks just like the Patriot’s crawl and general unability to recover, which I considered quite telling. Of course as you can see, I do not have the same burden of proof than with the Patriot drive. I was just looking out for the signs, and I had them already.
Instead I went to format it thinking I’d do an add-on test with the drive filled up until the full drop down 100MB/s, as it seemed to have the same breaking point (from the first test) than the Patriot drive, which I pegged to be around 25% of the capacity filled. Just to see how it would behave in subsequent uses after a cool down and with just a 100GB add-on...
First time formatting through regular OS option...
...bricked/went into RAW partition. Targeted Web search on this series reveals a tendency to various such failures reported by many users. No further testing – no curiosity about its Reads potential, and quite out of patience with such inconsistencies. Sent back as is for a full refund this time – had no difficulty at all obtaining it. Restablishing it from the RAW state would have been easy but was NOT compelling just to prove a point about a drive I had no intentions to keep anyway. Take a look at the FINAL WORD.
SK HYNIX X31 1024GB – ADVERTISED AT 1050MB/s READS & 1000MB/s WRITES
Bought because Amazon had it for the cheapest ever money for a 1TB drive, even vs. the cheapest “no-name” China brands – and obviously, it was one of the 3.1 drives with some of the best reviews and (whenever tested) sustained Writes speed out there.
CDM CHART
CDM ASSESSMENT – BACKWARDS INTERFACE/LESSER HARDWARE
The SK Hynix fully meets its claim from its native 3.1 interface looking at Peak Sequentials in CDM, and even exceeds them for Writes which keep real close to Reads speed. It does pretty good of the Real World tests there as well: 85% of Peak Sequentials overall. However, where at a glance most RND4K Reads results are par for the course with most USB 3.1 devices, there are some quirks with RND4K Real World Writes where they were faster on USB 3.0 than USB 3.1 on my PC, yet either way 25-45% slower than the other drives tested on my PC . However, it’s Read speeds there were consistently 10-20% faster than the other drives, especially on my laptop. Peak RND4K were overall relatively stronger as well, and again especially on my laptop. All in all, I prefer its strenght and consistency over lesser hardware – and one gen down – to whatever quirk/lesser RND4K Real World Writes ahead.
The drive also fully meets my rule of thumb about backwards compatibility: Peak Sequentials are 75% of the bandwidth overall; Real World Sequentials Reads stronger than the other drives for similar Writes, to 90% overall of the Peak Sequentials; Peak RND4K about 77.5% overall and Real World RND4K 94-99(130)% of USB 3.1. The SK Hynix does the best of it for all drives tested, counting with or without the quirk between PC and laptop. It was expected that, notwithstanding Sequentials that are firstly impacted by bandwidth limitation, RND4K Q32T1 would be where the slower interface speeds would carry most impact even if in their case the theoretical bandwidth “normalized” by rule of thumb is largely sufficient.
The impact of lesser hardware is more than ever nothing to worry about with the X31: Peak RND4K takes the worst hit with values ranging from 70-90% (which is still excellent) of the results from my PC, and everything else is 90-100(130)% with a heavy concentration at the 95%+ mark. For typical use the impact is really negligible here where the SK Hynix really shines through.
The tightening effect observed with all other drives going down one gen obviously holds true here but much more importantly carries quite spectacularly over lesser hardware, making the X31 a highly consistent and seamless as possible performer in all uses.
REAL USE SCENARIO
*Please note that this data set is the same exact one used with the Kingston drive.
Writes: No Fill...
...for a 841GB movie transfer: drop from around 750MB/s to around 200MB/s at 63% = some 530GB filled = more than even 2TB drives with pSLC cache would sustain at full speed. Then the X31 recovered to full speed around 76% = some 640GB filled. Kept around 750MB/s until the transfer was nearly done = 510MB/s average.
Writes: Filled up to 841GB
...for an additional 111GB movie transfer: around 750MB/s without any drop. So THIS drive DOES recover, even further filling it within 0.5% of its full capacity with a rather additional large data set.
From a curiosity standpoint, and while I did no sustained Reads test, a quick cut/paste test back to the source drive with some of the movies showed nearly 900MB/s speed. Even if it were to drop a bit after some 25% of its capacity like the Patriot drive did, and up to 530GB as per the SK Hynix behavior in sustained Writes, I had no doubt it’d still be satisfying. I rather thought that the Reads test we’ll see below would be more interesting to illustrate the capacities of this so far very promising drive.
Writes: Varied file types with LOTS of 1KB files on top of various stuff from 50KB to 4GB/No Fill
... for a 280GB transfer: about 13 minutes = 370MB/s average and no permanent drop if a lot of expected dips. To keep things fair and square, the Patriot could do about the same speed in this test – until it went into an unrecoverable 20MB/s crawl before half was done.
Reads: Varied file types with LOTS of 1KB files on top of various stuff from 50KB to 4GB/Filled 280GB
...copying them files back into a “Test Folder” on the source drive they came from. Some 600MB/s average with just the expected dips from the varied file sizes, but no drop. Of course that’s something the Patriot drive wouldn’t be able to pull even before it would inevitably slow down like it did in a fully Sequential Reads test.
No further testing. A clear winner and keeper that comes with a bit of extra space over most “1TB” drives and keeps full speed for sustained Writes longer than most “2TB” drives with pSLC cache can while comparing favorably in cost/capacity buying two of them – and recovering from a speed drop while in sustained use! Also brings the possibility of similar advantages in a thumbdrive (admittedly rather bulky) form with the T31.
FINAL CONSIDERATIONS: THE RULE OF THUMB & CACHE/FILL LIMITATIONS
My initial findings and extrapolating with my old Sandisk Extreme drive from USB 3.0 to USB 2.0 proved right with all three USB 3.1 drives tested – where CDM is concerned and from a USB 3.1 to USB 3.0 port in their case. The tightening effect proved right as well when going backwards one gen. Lesser hardware would not be a particularly noticeable drawback to any of the three drives tested in typical use, but especially the SK Hynix was a consistent drive in all tests that really shined when benchmarked even from a 3.0 port on my laptop – or in real use.
While the sample tested here is not representative of the whole market, two drives from different brands and form factor and components behaved relatively similarly in that, once filled to around 25% of their capacity they are permenantly crippled until emptied out below that point or formatted anew altogether. The Kingston was a bit more forthcoming than the Patriot, in that it could do some 25GB add-ons at best speeds from there (while the Patriot just about instantly went to a crawl), but then went way below what it could even hold in a sustained Writes test, almost exactly mirroring the Patriot’s crawling speed.
With the Patriot was also found that Reads would take a noticeable hit whenever reading back more than 25% of its capacity as well, but to a lesser extent than with Writes, and reasonably fast even there. However the point stands that if you need to Read as much you need to have filled any of them as much in the first place, so it circles back to the Writes problem.
The only notable difference between either the Patriot or the Kingston, and the SK Hynix which behaved splendidly in a sustained Writes test AND could recover in subsequent uses when filled some more even up to 99%+ of its capacity is that the Hynix has a DRAM cache while the other two only rely on a pSLC cache. While I do not have enough drives tested to make it into a general rule, nor did I test the Kingston thoroughly enough to make sure, I can still issue a very severe warning about drives without a DRAM cache especially, but also that it seems as much of a thing with capacity filled up to 25%+ than strictly a cache thing. I’d test any such drive at least thoroughly enough to make sure no such behavior is readily observable.
Therefore I’m not entirely sure what causes the effects I saw/where cache plays into it. A good exemple of this is just comparing the Kingston XS1000 1TB I tested vs. what TPU said in a review of the 2TB unit of the same Kingston series:
“Sustained write speeds of the Kingston SSD aren't that impressive. With just 180 MB/s to fill the whole disk, it's slower than many of its competitors, but considerably faster than the Crucial X8. If you plan on copying hundreds of gigabytes per session without pause, then look elsewhere. If your working sets are smaller, up to 100 GB writes, or so, or you have pauses in transfers that give the drive room to breathe, the XS1000 will run much faster, reaching around 650 MB/s—the pseudo-SLC cache size is 120 GB, which is a reasonable size.” https://www.techpowerup.com/review/kingston-xs1000/
Obviously TPU reviewers know more than I do, so on the one hand IF the pSLC cache of the Kingston 2TB is 120GB, surely the 1TB I tested has at most an equal cache. Yet, the 2TB tested by TPU would hit that 180MB/s wall in a sustained Writes test just about when the 120GB cache was out of the equation, whereas mine was perfectly fine until 185GB filled, and only got to its slowest average speed of 100MB/s after 235GB filled.
Then again, TPU remarked that the drive would run much faster with smaller working sets “up to 100GB or so” or with pauses to “give the drive room to breath” between transfers, but mine could only sustain about 25GB add-ons at full speed before dropping even lower than 100MB/s – the so-called crawl I’ve been talking about, of which I never heard a word in many reviews, even those that do a sustained Writes test, and also relay the hypothesis that the drive would regain full speed with smaller data batches or with breathing room between use. Reminder: the Patriot didn’t have nearly as much breathing room whatsoever into that crawl.
Of course, there might be good reasons why TPU got what they got and why it’s so different from what I got with quick testing the Kingston XS1000, and I’d sure like to have @W1zzard POV on my findings, but from where I stand I’d be particularly wary of bigger drives because they could very possibly be even more useless as per capacity than the ones I tested here. But more generally speaking, and reiterating the point, TPU is not the only source that do a sustained Writes test and note a drop equivalent to the cache size filled, nor the only source to maintain that the drive would likely recover with breathing room or smaller data sets.
Hope I have contributed enough for folks to make up their own mind about it. Hopefully, some will try similar experimenting with their drives and prove me wrong, perhaps observe different symptoms, or confirm in some cases that such a problem exists. More data around this would surely be appreciable.
Last edited by a moderator:
