Over at the Atola blog, Vitaliy Mokosiy published How to buy a reliable SSD – and none of it is wrong. That’s the problem. Read it back and it boils down to buy a good drive, don’t buy a bad one, and make backups. Every line is true and every line is the line an AI would hand you from a one-sentence prompt.

Coming from a reputable data recovery lab, that’s a missed opportunity. A lab sees the drives that actually died, so it should be able to tell you which good advice quietly stops being good, where the marketing hides the catch, and how you’d even check. So here’s the same checklist with the receipts attached – the version we’d expect from people who open these things up for a living. (We’ve done some of that homework ourselves: see Why SSDs Die a Sudden Death and Identifying SSD Controller and NAND Configuration.)

First, that Samsung 990 Pro article

The article leans on the 2023 Samsung 990 Pro “rapid health degradation” story as Exhibit A for SSDs die suddenly and without warning. There’s a lot wrong with using it that way.

One: the 990 Pro is a flagship TLC drive with DRAM, a five-year warranty, first-party firmware tooling (Samsung Magician), from about the most reputable consumer brand there is. It sails through every single recommendation the same article makes below. If the checklist can’t catch its own headline horror story, what is the checklist for?

Two: it’s a January 2023 article. Samsung shipped the fix – firmware 1B2QJXD7, pushed through Magician – in February 2023, and drives manufactured from around September 2023 onward left the factory already patched. Citing it in 2026 as a live cautionary tale is like warning people off a car over a recall that was closed three model years ago.

Three: it actively undercuts the “silent death” thesis. Dig into what the fix did and the picture flips – the accelerated “wear” was largely a SMART misreport: a firmware bug in the health calculation that ran the wear counter down far faster than the NAND was aging. Drives were flagging near-zero health after writing as little as a couple of terabytes against a 1,200 TBW rating. In other words the drive wasn’t dying quietly – it was screaming through SMART, loudly and early, and a firmware update stopped the bleeding. That’s the opposite of the story it’s being used to tell.

Four: it quietly proves the article’s own “firmware tools matter” point far better than the article argues it – and then stops one step short. Samsung isn’t special here. SK Hynix’s flagship Platinum P41 (and its Solidigm P44 Pro twin) shipped with a pseudo-SLC cache-clearing bug: under certain conditions the fast write cache stops flushing, writes fall through to raw TLC, and sustained speed roughly halves. Reports go back to 2022; a proper firmware fix (51061A20, via SK Hynix Drive Manager) only turned up years later. Same shape of problem, same fix mechanism, a flagship drive, and no mention. If firmware is the safety net, the reader deserves to know how long the net can take to show up.

The checklist

“Avoid QLC NAND memory. Prefer TLC instead.”

Yeah, right. And what about the manufacturers who switch between TLC and QLC batches interchangeably, or mid-production? What about models that ship TLC in the smaller capacities and QLC in the biggest, most desirable one – while the review samples happen to be the good version?

Example? The WD Blue SN5000 uses TLC on its 500GB/1TB/2TB SKUs and QLC on the 4TB – the capacity most people buying “a big drive” actually want. Reviewers were shipped the 4TB, and it was enough of a trap that PCWorld had to publicly correct its own review in August 2025 after first reporting the 4TB as TLC. The successor SN5100 then quietly went full-QLC across the entire line. Another one? Patriot Memory Viper VP4300 Lite, generally TLC, 4TB SKU QLC. If the outlets whose job is to catch this got it wrong, what chance does the checkout-page shopper have?

These details are buried deep in spec sheets, when they’re disclosed at all. I’ve been burned by exactly this, more than once. So the real question the advice skips: how do you tell what’s inside the drive in your hand? There are tools – we wrote up the Flash ID utilities back in 2019, which read the controller and NAND type straight off the chips. Those particular tools are SATA/USB-only and won’t touch NVMe, so for a modern M.2 drive you’re either using a modern version of the tools (SSD utils (20.06.2026) for Windows or the Linux equivalent) or leaning on teardown photos and reviewers who actually delid the thing. Point is, “prefer TLC” is the start of the job, not the whole of it.

“Avoid cheap brands – they use low-quality components and update firmware poorly.”

And which ones are those, exactly?

Is Kingston a cheap brand? Their budget NV1/NV2 line is a documented hardware lottery – one model number, different controllers (Phison E13T/E21 or Silicon Motion), TLC or QLC, always DRAM-less, with Kingston advertising only the sequential numbers any combination can hit. Buy one and you don’t know what’s in the box. Fine, you say — so pay up and buy the flagship instead. Except the premium KC3000 / Fury Renegade rolls the dice too: some units ship Micron’s 176-layer B47R TLC, others Kioxia’s older 112-layer BiCS5, and the only way to tell before you crack it open is the firmware string on the sticker – EIFK31.x is Micron, EIFK51.2 is Kioxia. Same drive, same price, same box, different silicon. And that’s the whole point: the expensive, DRAM-equipped, five-year, TLC-marketed pick this checklist would wave straight through still makes you gamble on what’s actually inside.

Is Western Digital/SanDisk a good brand? Also yes – and they got caught silently swapping the NAND in the popular WD Blue SN550 in mid-2021, quietly editing the datasheet while sustained write speed dropped from around 610 MB/s to about 390 once the cache filled. They only promised to use new model numbers after being called out. And they’re in good company: Adata, Crucial, Silicon Power and even Samsung (the 970 EVO Plus) have all shipped post-review component changes.

Is Samsung a good brand? You just quoted their flagship’s 2023 firmware fiasco to open the article.

So “avoid cheap brands” collapses on contact. The brands aren’t the unit of trust – specific models and specific production runs are. Telling someone to “buy reputable” without naming the traps is the advice equivalent of “invest in good stocks.”

“Consider the specified drive write endurance (TBW). Higher is better.”

Right. I did. And you know better than I do that TBW barely means anything for a consumer. It’s a warranty ceiling, not a reliability prediction – set conservatively, and one that the overwhelming majority of home and office drives will never come close to touching in their service life. If it’s useful at all, it’s as a tell: an oddly low TBW-per-terabyte is one hint you may be looking at QLC that the spec sheet didn’t bother to spell out. (The SN5000 again – the 4TB QLC’s endurance doesn’t scale the way the TLC capacities’ does.) Reading it as a ranking of “how long will this last” is reading it wrong.

“Check for DRAM cache – preferred for OS drives and sustained workloads.”

This is advice from the SATA era wearing modern clothes. DRAM on an SSD mainly holds the mapping table (the logical-to-physical translation layer). It mattered a lot on old SATA drives, and it still matters on weak, few-channel budget controllers – which struggle either way. But modern DRAM-less NVMe drives borrow a slice of system RAM through HMB (Host Memory Buffer) to do the same job, and for practically any real-world consumer workload the difference has shrunk to something you’ll measure but never feel. A good current DRAM-less NVMe drive will comfortably beat an older, lesser DRAM-equipped one.

DRAM-vs-no-DRAM only cleanly favors DRAM when everything else is equal – same generation, controller, NAND. It’s nice to have if you’ve money to spare (and at today’s prices that’s a real “if”), but it is not something to hunt for, and it is certainly not a reliability feature. Hunting for it while ignoring the controller and NAND generation is optimizing the wrong variable.

“Check warranty and support – choose 4 years or more.”

Four? The drives actually worth buying don’t ship with less than five. Samsung’s 9-series, WD Black, Kingston KC3000/Fury, Crucial’s performance line, SK Hynix Platinum – all five years. A three-year warranty is a signal you’ve wandered into the budget/lottery tier (hello again, NV2/NV3). What are those drives that come with a four-year warranty anyway?

“Availability of firmware tools – can you easily update the firmware?”

And how do you check that before you buy – from the product page? More to the point, the existence of an updater guarantees nothing. Plenty of models ship a first-party tool and then see zero firmware updates across their entire life. And plenty of models with no standing tool got a one-off updater bundled the moment a real bug forced the issue. What you want isn’t “is there a tool,” it’s “does this maker actually ship fixes when something breaks” – and the only way to know that is the maker’s track record, which brings us right back to reading the history rather than the spec sheet.

“If you can afford it, opt for enterprise-grade SSDs – optimized for endurance and reliability.”

Sure – if you have enterprise-grade workloads. For a consumer? You’d be buying power-loss-protection capacitors, U.2/U.3 form factors and a sustained-write profile tuned for a server rack you don’t own. And here’s the kicker that undercuts this sitting right next to “avoid QLC”: a large share of modern enterprise drives are QLC – the big high-capacity datacenter models especially (Solidigm’s D5 QLC line runs to tens of terabytes per drive on QLC). “Avoid QLC, but also buy enterprise” is contradictory advice.

“For PCIe 4.0/5.0 or M.2 2230 drives, think about airflow or a heatsink.”

How, though? Is the motherboard’s bundled heatsink enough? Should you buy drives with heatsinks pre-attached, or skip those and bolt on a third-party block? Does it even matter?

Mostly it affects speed, not survival. The hot component is the controller, and when it heats up the drive throttles – deliberately, to protect itself – which costs you performance, not lifespan. Sustained catastrophic overheating is a different story, but if you’re hitting that, your whole system’s cooling needs help, not just the SSD. And if it’s a laptop? You’re not adding a heatsink anyway, so be real about the advice. Telling someone to “think about airflow” without saying what it does or what “enough” looks like just adds anxiety.

“Gather feedback – Reddit, HDDGuru, Tom’s Hardware, TechPowerUp, store reviews, PCPartPicker.”

Sound advice, at last – with one blind spot. It’s close to useless for drives that are new on the market, which are exactly the ones where the bait-and-switch risk is highest and independent teardowns don’t exist yet. The freshest launch is the one you can research the least.

“Always back up important data.”

Do everything good and don’t do anything bad. This is where the advice stops being generic and becomes a truism. Everyone reading it is already in one of two camps: they back up, or they haven’t lost data yet. Nobody moves from the second camp to the first because a blog post told them to – they move because it finally happened to them. (For the record, even Samsung’s own fix flow told 990 Pro owners to image the drive before flashing. Good advice; also table stakes.)

“If you buy in bulk, diversify – controller, NAND generation, production window, firmware.”

This is the one genuinely non-obvious idea in the piece – the correlated-failure insight a recovery lab really does own, and the reason a rack of identical drives can go down together, as that HDDGuru thread showed. Credit where it’s due.

The practical catch: diversifying means multiplying the spec-sheet detective work above across every model you’re comparing, and if you’re buying at that scale you’re likely not the person who gets to overrule the purchasing department’s single-SKU, single-vendor contract. Great principle; harder to actually deploy than a bullet point makes it sound.

“And, surely, ask AI about the model and failure cases. Deep research is highly recommended.”

Now that’s a good idea – with one condition that makes or breaks it. Use a model that will actually search the live web, not one answering from its own weights. Ask an offline model about a drive and you’ll get confident, fluent, out-of-date answers – the firmware fix it never heard about, the QLC swap that happened after its training cutoff, the failure thread from last month it can’t see. On a topic where the whole game is which revision, which firmware, which production window, stale knowledge isn’t just unhelpful, it’s actively misleading.

Which, fittingly, is the same lesson as the rest of this list. The advice isn’t wrong. It’s just what you get before anyone does the homework.