Re: SSD TBW

On 8/21/2026 12:32 PM, Scott Dorsey wrote:
>   =?UTF-8?B?IEdsb3N0w6lpcg==?=  <[email protected]> wrote:
>>
>> I do not yet have a first-hand experience of an SSD failure. Don Y
>> complains in news:sci.electronics.design that an SSD suddenly
>> completely failed, instead of gradual partial degradations of a hard
>> disk. What does Lawrence D’Oliveiro mean by
>> "before
>> showing signs of failure."?
> 
> With a conventional hard drive, first you might see some error numbers
> creeping up on your SMART tool, as individual blocks start failing.  There
> is a bad block replacement process that is transparent to the OS so all you
> see are bad numbers on SMART.  Then you start getting media errors and
> then it's all over.

The firmware in spinning rust controllers has matured over decades.
This is not the case with SSDs (especially early entrants to the market).

Note that the media fails in different ways and the controller has
to understand these and adequately address them, in the wild.

Many "Thumb drives" have had problematic firmware (some Phison and
Hynix controllers).  The fix requires a firmware upgrade -- WHILE
the drive is still willing to talk to its host.

>> This quotation gave me the impression that Lawrence D’Oliveiro does
>> not expect the dramatic all-or-nothing scenario that Don Y reported.
> 
> You can also have dramatic all-at-once errors when the interface fails.  I have
> seen many USB SSDs where the SSD memory remained fine but the USB interface could
> not get to it.

Or, it suddenly becomes R/O -- better than inaccessible but only if
you're not in the process of trying to write to it!

> I have seen lots of ways that SSDs fail.  There are likely lots more than I
> haven't seen yet.
You can likely run a magnetic disk for many years (I have drives with 80K PoH)
that still haven't encountered a remapped sector.  But, you can exhaust the TBW
limit for an SSD in a short time -- if you are ignorant of this limitation.
At 200+MB/s, you can scribble 12GB in a minute -- almost a TB in an hour!
The type of FLASH used, extent of overprovisioning, level of smarts in
the controller, etc. all have a big impact on real-world numbers.

In the early days of WAROM (e.g., ER3400), naive implementations that had
previously used BBSRAM for that nonvolatile function would wear out in
minutes ("No, you DON'T want to write the changed settings back to the
store each time an individual setting is changed!  Keep a shadow copy in
RAM and use an "impending power fail" signal to quickly stash them to the
medium *IFF* (!) SOMETHING HAS CHANGED.  It doesn't take long to go through
10^4 erase/write cycles!