13

Modern PCs tend to use one of two types of internal-storage devices:1 2

  • Hard-disk drives (HDDs) store data on spinning disks coated in a magnetic recording medium, a technology dating back to the 1960s. HDDs are very cheap and available in enormously-high capacities, but relatively slow and mechanically quite fragile (they have lots of parts moving really really fast that have to be kept in extreme close proximity to one another without colliding).
  • Solid-state drives (SSDs) store data in semiconductor memory, either flash EEPROM or (occasionally) battery-backed DRAM or SRAM. SSDs are faster than HDDs and far more mechanically robust due to their lack of fragile moving parts, but are also much more expensive per byte, not available in capacities nearly as large, and less reliable for long-term storage (as their memory cells tend to slowly leak electrons over time).

Both types of storage device wear out over time and eventually fail. However, it seems to be common knowledge that SSDs always, or almost always, fail catastrophically, all at once, with little or no advance warning; in contrast, HDD failure is usually gradual, with a long period of slowly-declining performance and increasingly-frequent occasional fsck/chkdsk-warranting read/write errors, providing plenty of time to back up the contents of the drive and swap it out for a fresh specimen.

While both HDDs and SSDs have obviously-catastrophic failure modes (such as a failure of the drive controller or interface), one would expect, for both types of drive, that most drive-wearing-out failures, at least, would occur gradually, in the manner classically associated with HDDS, with the drive gradually becoming slower, encountering increasingly-frequent read/write errors, and accumulating bad sectors (for HDDs) or memory cells (for SSDs) over time, until the degradation finally corrupts something vital to the operation of the drive.

Both types of drive deal with worn-out portions of medium in the same manner, by silently remapping data from dead sectors or cells to an internal store of spare sectors/cells set aside at the factory for this purpose;3 even once this stash of spares has been exhausted, and bad sectors or memory cells start becoming visible to the OS using the drive, the filesystem used on the drive still provides a second layer of protection by remapping from these bad portions of storage medium to unused, still-good storage.

Indeed, one might naively expect that HDDs would be somewhat more prone to sudden, catastrophic failure than SSDs, for a couple of reasons:

  • HDDs possess catastrophic failure modes not present with SSDs (such as a failed or seized spindle motor or a head-disk collision).
  • With HDDs, the lowest level of granularity is the sector (a chunk of drive 512 bytes in size, or 4 KiB on some newer drives), whereas SSDs work with individual memory cells (storing just one to four bits each), allowing much-finer-grained handling of good and bad sections of the storage medium and permitting more efficient use of what good space remains on the drive.

Yet, despite all this, HDDs usually fail gradually, while SSDs usually fail all at once. Why? Could this be related to how most SSDs start redistributing writes away from highly-used memory cells even before they start to fail, delaying the first few outright cell failures (but, even then, one would expect there to still be considerable advance warning of an impending drive failure, as individual cell failures are statistical rather than deterministic, making it impossible to eliminate the early-warning cell-failure tail)?


1: Obviously, a computer is not restricted to using just one or the other; as long as it has enough connectors to connect two or more internal drives, it is perfectly possible to install both an HDD and an SSD in the same computer alongside one another. This can be used to exploit the advantages of both types of storage device, with the smaller, faster SSD holding the operating system and disk-access-limited applications, and the bigger-but-slower HDD holding bulk data and non-disk-access-limited applications.

2: Some newer drives use both methods of data storage, combining a large, slow magnetic drive with a smaller built-in solid-state cache; these work on the same principles as the setup described in footnote 1, but generally appear to the operating system as a single drive, rather than two, and do not usually offer the user the option to specify whether files should be written to the magnetic storage or to the solid-state memory. Such drives are commonly known as fusion drives, not to be confused with the method of starship propulsion by the same name.

3: This is why modern HDDs and SSDs generally report as having no bad sectors or cells at all when new, despite the imperfections inherent in drive manufacturing, and why bad sectors/cells only start becoming visible to the OS late in the drive's lifetime, once the drive's internal stock of spares has been exhausted.

2
  • 6
    I'd challenge your premise, as exclusive. Many HDs fail with the 'click of death' rather than slow decline. Many SSDs suffer SMART-noticeable 'slow' fail. Each requires adequate backup strategy & costs a new drive. Whether that is over days or weeks, that's what backups are for.
    – Tetsujin
    Dec 19, 2021 at 18:40
  • 1
    Internally, the granularity of wear leveling and good-vs-bad areas in SSDs is limited by addressing rather large swaths of individual memory cells, grouped into pages and blocks. "The size of a NAND-flash page size can vary, and most drive have pages of size 2 KB, 4 KB, 8 KB or 16 KB. Most SSDs have blocks of 128 or 256 pages, which means that the size of a block can vary between 256 KB and 4 MB." <codecapsule.com/2014/02/12/…>
    – Nate
    Dec 30, 2021 at 6:54

3 Answers 3

12

The typical reason for SSD "sudden-death" can be attributed to the functional requirements for their operation. @LawrenceC touched on this briefly in their answer, but I'll expand on it here.

All SSDs require certain metadata to function:

  • Controller firmware to establish baseline functionality
  • Microcode updates to alter functionality or encryption keys
  • Wear-leveling data used by the Flash Translation Layer
  • S.M.A.R.T. attributes, etc

All of the above are stored in the so-called system area of the NAND flash. Roman Morozov, a data recovery specialist at ACELab, provides helpful details on an excellent blog post at Elcomsoft (emphasis are mine):

The system area contains SSD firmware (the microcode to boot the controller) and system structures. The size of the system area is in the range of 4 to 12 GB. In this area, the SSD controller stores system structures called “modules”. Modules contain essential data such as translation tables, parts of microcode that deal with the media encryption key, SMART attributes and so on.

The translation table—ironically, the very mechanism which enables wear-leveling—cannot be wear-leveled:

If you have read our previous article, you are aware of the fact that SSD drives actively remap addresses of logical blocks, pointing the same logical address to various physical NAND cells in order to level wear and boost write speeds. Unfortunately, in most (all?) SSD drives the physical location of the system area must remain constant. It cannot be remapped; wear leveling is not applicable to at least some modules in the system area. This in turn means that a constant flow of individual write operations, each modifying the content of the translation table, will write into the same physical NAND cells over and over again...

Since the NAND System Area cannot be wear-leveled, it experiences much higher stress than the data area. This is exacerbated by frequent small writes:

Such usage scenarios will cause premature wear on the system area without any meaningful indication in any SMART parameters. As a result, a perfectly healthy SSD with 98-99% of remaining lifespan can suddenly disappear from the system. At this point, the SSD controller cannot perform successful ECC corrections of essential information stored in the system area. The SSD disappears from the computer’s BIOS or appears as empty/uninitialized/unformatted media.

This can manifest itself as completely dead drive if the controller can't boot the firmware:

If the SSD drive does not appear in the computer’s BIOS, it may mean its controller is in a bootloop. Internally, the following cyclic process occurs. The controller attempts to load microcode from NAND chips into the controller’s RAM; an error occurs; the controller retries; an error occurs; etc.

But more often than not, the data will simply disappear because the FTL has failed:

However, the most frequent point of failure are errors in the translation module that maps physical blocks to logical addresses. If this error occurs, the SSD will be recognized as a device in the computer’s BIOS. However, the user will be unable to access information; the SSD will appear as uninitialized (raw) media, or will advertise a significantly smaller storage capacity (e.g. 2MB instead of the real capacity of 960GB).

So the sudden failure of SSDs has always been a result of limited write endurance—but of the drive metadata, not the user data.

1
  • 4
    Why couldn't the system area of a drive support wear leveling by having a few reserved areas, along with one area that identifies which of the other areas is currently being used? Even if the last area couldn't be "wear leveled", it would only need to be written once for every 10,000 times the next layer was written.
    – supercat
    Oct 26, 2022 at 22:46
6

You have most of the points correct, but here's a few that possibly explains your question:

SSDs are only generally faster than HDDs. The largest capacity hard drives actually rival the low and mid range SSDs in speed. Hard drive bandwidth tends to be proportional to data density. Write speed on SSDs can actually be twice as slow as read speeds.

SSDs and HDDs have very different wear issues and failure modes. In general, HDD wear is from start up and stop, and continuous hours spinning, and their wear is only statistically predictable -- some drives can easily last 5-10x their warranty period. The spinning itself doesn't kill the HDD, but some random wear event influenced by the spinning that may or may not occur. Most current SSDs have a limited number of write cycles per block, and when they reach that count, they will fail very soon after. Enterprise grade SSDs are rated in max total writes per day over the warranty period.

Both SSDs and HDDs include an internal failure prediction system. In linux, the smartmontools with the smartctl utility can read the SMART data from most of both of these drive types. Some SSDs instead need the nvme-cli package which extracts similar data. Similar tools should be available in windows, and some disk vendors also have their own tools. Using these tools, you can usually detect impending failure long before the drives fail. Some SSDs actually will tell you their exact health (as a percentage of write cycles used so far). In those cases, a new SSD will tell you how soon it will fail. With a HDD, it can't tell you when it will fail, it can only tell you if a random failure event has already occurred and the drive is already failing.

In my experience, hard drives sometimes give a week or two of warning before catastrophic failure, and this comes in the form of increasing occurrence of uncorrectable I/O errors. Typically, a flake from a wear spot floats around inside and scratches the rest of the drive, and it fails fairly suddenly with very little warning before data loss has already started. Spindle failure in disks is rare, and probably a manufacturing defect.

Most SSDs use something called wear leveling, which reuses blocks that haven't been written in a while by copying their data to another block and then writing new data to it, so that all blocks get the same number of writes. This extends the life of the drive, but the result is that you get no errors until it does fail and it fails evenly and all at once. It doesn't get "wear spots". But the drive knows how close it is to failing from the very beginning, and if you ask it, it will tell you.

As already pointed out, if you drop a HDD, that may kill it instantly. So with all this, to say an SSD fails instantly where a HDD warns you is a bit backwards. Of course, either one can fail suddenly with a controller failure.

SSDs do not work with individual bits. They work in blocks, typically 2048 bytes, and when you write a block, it has to find an unused block or it has to relocate an under used block and then replace it. SSDs are actually less fine grained that disks, so you've got that backwards too. But really block size is an implementation detail, and has gotten bigger over the years for disks too.

8
  • 3
    Re: your second paragraph. What actually makes SSDs faster is not their transfer speed but their access time (and consequently IOPS for random access), which is orders of magnitude faster than rotating HDDs.
    – StarCat
    Dec 21, 2021 at 15:32
  • 1
    Bandwidth is important. Access time and iops can be covered up with cache and buffering, both in the OS and the drive, but that doesn't fix latency, so all of this factors in.
    – user10489
    Dec 21, 2021 at 15:48
  • 1
    Also, higher capacity disks have more tightly spaced tracks too, so head latency is lower track to track to some extent. It's still the same distance (and time) from first to last track through.
    – user10489
    Dec 22, 2021 at 14:47
  • 1
    Apparently it isn't all that non-deterministic, especially when averaged over all the cells.
    – user10489
    Dec 30, 2021 at 12:57
  • 1
    "But the drive knows how close it is to failing from the very beginning, and if you ask it, it will tell you." - It seems like the OS should be asking it regularly, then, and generating some sort of useful warning when it's getting close to the end of its lifespan. I know there are tools to do this, but given that most modern computers now contain an SSD, and Aunt Tillie knows nothing about such tools until her computer suddenly dies without warning, it would be nice if the OS did this on its own. Apr 28, 2022 at 20:40
3

Since the advent of IDE drives, and likely even before, each hard drive is basically an embedded computer platform--it has a CPU, RAM, runs firmware, the firmware talks to the PC over a bus/port and also talks to the device electronics over internal I/O interfaces. The firmware likely either uses the storage medium itself to store things or additional flash on the circuit board.

Unless you are somehow "inside" of this embedded computer, there's really no ability to tell the exact reason why a hard drive has failed, unless you have obvious external signs like noises/platter scratching sounds. For SSDs you don't have these. SMART gives you some abstracted information and clues but nothing like a "crash dump" to analyze at the moment of failure.

Now there are likely hidden/undocumented service interfaces on devices used during manufacturing and servicing, but without a lot of reverse engineering work, no one's going to know about them and how to use them, and these techniques likely vary greatly depending on manufacturer, model, series, etc.

Yet, despite all this, HDDs usually fail gradually, while SSDs usually fail all at once.

Given the above, honestly we can really only speculate. Some information for speculation:

  • NAND works vastly different than platters. NAND needs to be erased before being used, erase-blocks are larger than data blocks, and writes wear out the flash. None of this applies to platters.

  • SSDs attached to the same interfaces that HDDs work with could not be honest about how they worked, they needed to look like platter HDDs.

  • So both of the above mean that SSD firmware from the outset has to be more complex. More complex = more opportunities for failure.

  • If device firmware runs into a bad condition while booting and can't boot, or simply refuses to boot, the result is a dead device. Overall the chance for this increases due to the additional complexity we infer from the additional tasks we know SSD firmware is doing.

  • Possible exact reasons for this are myriad and you have no further visibility for a specific failure unless there is access to device-level diagnostic/service interfaces.

  • Ideally an SSD with no further ability to write would simply become read only. Unfortunately there are no standards here and nothing stopping an SSD manufacturer from putting firmware out there that simply gives up and refuses to boot if it's out of spare flash space to support additional writes.

  • There are "legitimate" things that can happen where the firmware really doesn't have much of a choice except to "die" - e.g. if an entire flash chip just died or stopped responding (most SSDs have many of these), if the SSDs internal RAM became bad, if various fixed locations on internal flash the SSD uses to boot or do basic tasks dies, etc. Do these things happen more often than the equivalents with platter drives? Who knows. I wouldn't even know where to go to get that data.

I've heard of some SSDs that have UARTs on what used to be the jumper pins on classic HDDs. Perhaps when the devices don't boot, something is output there.

But overall, the main point is, with most devices, we simply have no way to tell why SSDs suddenly stop booting, and given that SSD firmware has to be more complex and time-to-market is everything with new technology, we shouldn't be surprised that total failure happens more often.

1
  • 2
    If a drive supports SMART (or the nvme equivalent), you can directly ask it for its health information and get a pretty good idea of if, how, and why it might be failing.
    – user10489
    Apr 29, 2022 at 22:25

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .