SSD endurance test

Just over 18 months ago, Tech Report began a journey to do one simple thing – drive a set of six SSDs to their deaths and analyse the health of the flash memory in stages as things went along. The goal was to see how long it would take a user with a slightly abnormal workload to go through enough write cycles for the drives to eat through their flash reserves and eventually brick themselves. The results were both enlightening and spectacular, with every drive lasting through, and preserving performance, well beyond their rated lifespans set in their warranties. Not only is flash memory storage extremely reliable today, it lasts far longer than the best spinning disk we’ve ever produced for similar workloads. Hit the jump for a summary of the findings of the experiment.

All silicon can and must die

Its worth noting that flash memory operates on things called write cycles. Basically, the memory chips store electrons inside them in partitions called cells, and in specific places these electrons can be positioned to represent the numbers 0 or 1, bits that make up binary numbers. Writing to a flash chip sends electrons through it which get stored in the memory cells. Reading that flash chip sends electricity through a different channel at low voltage. Erasing a cell, effectively deleting the data, requires you to flood the cell with electricity, removing any electrons trapped inside. The more you use your drive, the faster the memory inside wears out and breaks down.

Its the breaking down part that is crucial to the story here. All silicon will die, because eventually it is no longer able to hold electricity in the transistors, in the traces on the PCB, that the silicon begins to degrade and break down over time. This process is sometimes fast and sometimes slow, depending on how much voltage you’re pumping through it. Its the reason why a CPU that is overclocked and overvolted has a much shorter lifespan than the same chip overclocked and undervolted for energy efficiency.

The silicon degradation process is also quite predictable, so SSD manufacturers usually provision a part of the drive as a reserve to replace any flash memory that dies out as a result of normal wear over time. This is why some drives are marketed as 240GB models instead of 256GB – the manufacturer sets aside 16GB of storage in its reserves to be used before the controller empties the pool completely and bricks itself to avoid users using it further, resulting in catastrophic data loss.

How it all failed spectacularly

ssd endurance test earlyfailures

When it came to the drives bricking themselves completely, four of the participants in the test failed to reach 1.2 petabytes (1.2PB) of writes before keeling over. Intel’s 335 was by far the most abrupt by design – once the drive’s internal mechanisms determine that the flash memory cannot be trusted (even with just one reallocated sector), you’ll be greeted one day with a boot error saying that the drive is kaput and you must proceed to immediately back up your data. If you don’t, the drive is guaranteed to brick itself on the next POST. What fun.

The other drives died more predictably. Samsung’s 840 lasted to almost 1PB of writes before bricking itself, generating predictable errors along the way that it was reallocating flash storage from its reserves. This particular drive is based on TLC flash memory, where the silicon stores three electrons per cell, equitable to three bits of binary code. TLC flash wears out much faster than MLC (two bits per cell) or SLC (1 bit per cell), but it is much cheaper than anything else. Next on the chopping block was the Kingston HyperX 3K 240GB, which wore out all of its flash reserves between 600 and 700TB of writes, bricking itself after a reboot just like Intel’s 335.

Taking the most extreme example of the Intel 335’s lifespan of 700TB with 20GB of writes per day, the drive would last you a record 35,000 days, or 95.8 years, before dying in the same fashion due to silicon wear. 

Corsair’s Neutron GTX 240GB delivered astounding performance, with perfect flash attributes all the way up to 1.1PB before failing sharply and predictably all the way to 1.2PB and almost 3400 reallocated sectors. Because the Neutron GTX used MLC flash, that’s almost 6.8GB of data held in the reserves before things became too hot to handle. In all instances seen here, the drives would far outlast their rated warranties. If you used them as one normally would with an average office-level workload, you’d only write 20GB of new data every day thanks to things like website caching, generation of log files and so forth.

Taking the most extreme example of the Intel 335’s lifespan of 700TB with 20GB of writes per day, the drive would last you a record 35,000 days, or 95.8 years, before dying in the same fashion due to silicon wear.

The final standings

ssd endurance test hyperx-writes

Like the Intel 335, the second Kingston Hyper-X 3K had an abrupt death, but had a regular, predictable failure pattern with the drive constantly pushing out errors that it was nearing the end of its reserves. Thanks to the compression used by the LSI Sandforce controller, any data that was compressible benefited the drive’s lifespan tremendously, running through 2.1PB of writes, but only actually writing arounf 1.5PB of data to the flash itself. That is a huge change in comparison to the other Hyper-X 3K which only had a diet of incompressible data, suffering a failure at just over 800TB. The Sandforce controller is the only one on the market that does offer write compression, so any current SSDs based on Sandforce hardware, like the recently-announced Mushkin Eco², could have a similar lifespan if the manufacturer has taken the time to tune the controller to the particular behaviour of the flash memory it will be tied to.

At just over 2.1PB of writes, Tech Report’s Geoff Gasior had to write an average of 116.66TB of data every month for 18 months before getting the drive to kick the bucket. Just how many of you readers are going to write over your drive 486 times over the course of the month? None, I’m guessing.

ssd endurance test 840pro-life

The Samsung 840 Pro used the same controller as the 840 Evo, but Samsung paired it to MLC flash instead of TLC, changing the drive’s behaviour and performance hugely. Over time, the drive failed rather predictably, but the way that Samsung’s controller reports failures is a little buggy. As soon as the controller reaches a low point in the wear leveling count, it will stay at that level until it finally dies. This means that its impossible to know how degraded the flash actually is and even the used block reserves only saw significant dips starting at 800TB of writes.

Unless you’re paying attention to both atttributes, it would be very easy for someone to look at the wear leveling count and pronounce the drive as being suspect, even though it would go on to write over 2.4PB of data (take note of that, because I’ve seen plenty computer nerds declare their drives as being suspect far too early in their life). Looking at how predictable the decline is, the flash could have easily lasted up to 3.0PB of writes before being completely useless for any storage use. That is incredible endurance. At 2.4PB, you could write a full terabyte a day to the drive and it would still last you well over six years. Writing just half of that extends the drive’s theorised life to 13 years, outlasting any other component you’ve replaced in your machine at that point due to performance upgrades.

Thus, the question is now answered – SSDs are reliable! Tech Report’s testing was incredibly thorough, even going as far as to write a full set of data to the drive, unplug them for two weeks and then do error checking on them to see if the flash memory was holding the data properly. You can read more about the test results and the performance statistics of the drives in the source link below, but I think this finally settles the debate over whether to fit a system with a SSD or a hard drive – if the price isn’t keeping you back from doing it, a SSD is preferable over a hard drive for use as your boot and system drive every single time over a regular spinning disk.

No longer should anyone have any reservations about using SSDs – it now needs to be the gold standard for computers moving forward. I hope that laptop manufacturers are paying attention to experiments like this one and others because if there’s any one market that can push forward the adoption of flash storage for computers, it is the mobile market.

Source: Tech Report