There are still a lot of write-once-read-many workloads that can take advantage of the speed benefits that QLC SSDs bring over traditional hard drives without worrying about write endurance of the NAND. Despite the back-of-the-napkin low P/E rating calculations Micron still warranties all capacities for 5 years. They're clearly confident that the endurance will not be an issue as long as these are used for the intended use case. It's also important to remember that "unused" free space is still used for wear leveling so large drives like these will have even more free NAND available to help prevent individual cells from degrading.
It should, barring a serious firmware bug. SSDs should go read-only when they exhaust their write endurance, so that the data can be recovered. Enterprise SSDs are expected to retain the data for at least three months after reaching this point, and the standard for consumer SSDs is a full year of retention at end of life. Some drives have been observed turning into an effective brick if you power cycle them after they go read-only, but even then the manufacturer has the ability to query the drive for surviving data, which includes the SMART indicators.
Cells have 1500 P/E cycles but write amplification can reduce the usable number of drive writes down to 92 (7.68TB model: worst case 4KB random write causes 64KB to be erased and re-written = 16x write amplification = 0.8/0.05). The larger the model, the larger the smallest write block can actually be to the underlying arrays. Smaller drives have smaller blocks for these writes so random writes have smaller penalties (1.92TB model: worst case 4KB random write causes 16KB to be erased and re-written = 4x write amplification = 0.8/0.2).
Yeah, they're very low. But keep in mind that enterprise SSDs don't use SLC write caches, so the low-OP/low-endurance TLC drives are also pretty slow for random writes. The 8TB 5200 ECO (TLC) gets 9.5k IOPS compared to the 8TB 5210 ION's 4.5k IOPS. And even the 5210 is still more than 10x faster than a hard drive for random writes.
Not compared to the drives they are replacing. These would be used in a huge array which will benefit from multi link connections and the low latency of ssd will make a huge difference. IOPS numbers by them selves are meaningless.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
13 Comments
Back to Article
PeachNCream - Thursday, November 8, 2018 - link
Back of the napkin calculations put individual memory cell P/E cycles at about 350-600. Meh.WithoutWeakness - Thursday, November 8, 2018 - link
There are still a lot of write-once-read-many workloads that can take advantage of the speed benefits that QLC SSDs bring over traditional hard drives without worrying about write endurance of the NAND. Despite the back-of-the-napkin low P/E rating calculations Micron still warranties all capacities for 5 years. They're clearly confident that the endurance will not be an issue as long as these are used for the intended use case. It's also important to remember that "unused" free space is still used for wear leveling so large drives like these will have even more free NAND available to help prevent individual cells from degrading.FunBunny2 - Thursday, November 8, 2018 - link
"They're clearly confident that the endurance will not be an issue as long as these are used for the intended use case."and when one croaks, what evidence will Micron demand that drive was used per "intended use case"? sounds like a Catch-22 situation.
Billy Tallis - Thursday, November 8, 2018 - link
The drive's SMART indicators give you plenty of fair warning if your workload is causing high write amplification leading to early wearout.FunBunny2 - Thursday, November 8, 2018 - link
but... does the SMART data survive the croak? real question, I've no idea.Billy Tallis - Thursday, November 8, 2018 - link
It should, barring a serious firmware bug. SSDs should go read-only when they exhaust their write endurance, so that the data can be recovered. Enterprise SSDs are expected to retain the data for at least three months after reaching this point, and the standard for consumer SSDs is a full year of retention at end of life. Some drives have been observed turning into an effective brick if you power cycle them after they go read-only, but even then the manufacturer has the ability to query the drive for surviving data, which includes the SMART indicators.PeachNCream - Thursday, November 8, 2018 - link
The 5 year warranty is indeed a good thing so there is that going for it.tygrus - Friday, November 9, 2018 - link
Cells have 1500 P/E cycles but write amplification can reduce the usable number of drive writes down to 92 (7.68TB model: worst case 4KB random write causes 64KB to be erased and re-written = 16x write amplification = 0.8/0.05). The larger the model, the larger the smallest write block can actually be to the underlying arrays.Smaller drives have smaller blocks for these writes so random writes have smaller penalties (1.92TB model: worst case 4KB random write causes 16KB to be erased and re-written = 4x write amplification = 0.8/0.2).
stanleyipkiss - Thursday, November 8, 2018 - link
Sell me a 8TB SATA SSD at less than $400 and I'll buy it. DWPD be damned.bubblyboo - Thursday, November 8, 2018 - link
Those write IOPS are downright frighteningtheeldest - Thursday, November 8, 2018 - link
If you're concerned with 4kB Random Writes, you shouldn't buy this drive.Billy Tallis - Thursday, November 8, 2018 - link
Yeah, they're very low. But keep in mind that enterprise SSDs don't use SLC write caches, so the low-OP/low-endurance TLC drives are also pretty slow for random writes. The 8TB 5200 ECO (TLC) gets 9.5k IOPS compared to the 8TB 5210 ION's 4.5k IOPS. And even the 5210 is still more than 10x faster than a hard drive for random writes.Dug - Tuesday, November 13, 2018 - link
Not compared to the drives they are replacing. These would be used in a huge array which will benefit from multi link connections and the low latency of ssd will make a huge difference. IOPS numbers by them selves are meaningless.