At 24 GB/day, or 8.8 TB/year, it would take 114 years to reach 1 PB.
Several SSDs have gone past that point, and still work fine.

Last year, I wrote about estimates for durability of SSD drives. Looking at how long it would take to exhaust the theoretical limit of the storage cells, based on different write speeds and quantities. The assertion was: It would take a very long time, and other technical failures were more likely to render the drive useless long before all cells were used up. In fact, the owner is not likely to outlive full cell exhaustion. This has now been confirmed through an experiment.

Over at Tech Report, they have done an endurance experiment lasting more than a year, with several consumer drives. For months, they have been writing data to the drives, while monitoring drive health, and verifying correctness. Some of the drives were rated for 20 GB / day for three days, which means about 22 TB. However, several of them made it past the 1000 TB, or 1 PB, mark. That’s about 50 times the advertised endurance rating. And two of the drives, a Samsung 840 Pro 256GB, and a Kingston HyperX 3K 240GB, have made to past 1.5 PB. It’s not all clear from the article exactly when they started, or how much they write per day. However, assuming a year to get to 1.5 PB, that’s 4.1 TB / day, or 47 MByte/s. The actual write speed is probably higher, but this leaves time for verification as well.

To put these numbers into perspective, I’ve extended the table from last time, and added a line for the speed of the Tech Report endurance test, as well as extra columns for multiple years total. As can be seen, the endurance test is running at 171 the write speed of what was identified as “heavy use”. Furthermore, the heavy use scenario is within what the typical consumer drives are rated for, i.e. about 20 GB/day.

MBit/s MByte/s MByte/hour GByte/day GByte/year TByte 3 years TByte 5 years
SATA3 max speed 6000 750 2700000 64800 23652000 70956 118260
Stress test 2000 250 900000 21600 7884000 23652 39420
Endurance test 376 47 171000 4109 1500000 4500 7500
Heavy use 2.2222 0.2778 1000 24 8760 26.280 43.800
Low/Medium use 0.0926 0.0116 41.67 1 365 1.095 1.825

What about the time to failure estimates; how do they compare to the empirical evidence? Last year, I noted that a 256 GB drive with cells of 10k write cycle life span (typical MLC memory), would take 256 years to reach 10% of failed cells, and about 350 years for full exhaustion, assuming 24 GB/day written. It turns out that was a bit optimistic. At 24 GB/day, or 8.76 TB/year, it would take “only” 172 years to reach 1.5 PB (where the Tech Report drives are now). If we go by 1 PB, it would take 114 years.

This is all for MLC memory. If we look at TLC, typically rated for 1000 write cycles, the endurance numbers are also a tenth. I.e. they would fail between 11 to 17 years of sustained 24 GB/day writes, again assuming a 256 GB drive.

If larger drives are used, the time to failure also increases. This is because there are more total space to level the writes across. In fact, doubling the size of the drive, will in theory double its lifespan. So an MLC 512 GB drive would last some 228 years, and a 1 TB drive 456 years. For TLC, the numbers are again a tenth, so 22 and 45 years respectively.

This is all well and good, and should at least put the final nail in the coffin regarding worries for SSD reliability. The only concern which the Tech Report experiment raises, is the way the drives fail when they do reach end of life. Of course there are plenty of relocated sector warnings in the SMART data beforehand. However, once they are past the point of no recovery, all data is lost. Several of them cannot be accessed at all. This is of course a bit different from spinning disks, which usually keep on reading some of the tracks, even if other parts are broken.

It highlights the fact that monitoring SMART data should be a standard procedure, and part of good data hygiene. Of course, a good backup strategy is required, regardless of drive type or usage pattern.