That Didn't Last Long: Samsung 960 EVO NVMe Already Fails

[microcode]

Samsung SSD RMA is awful. You call up the number and they tell you they sold the division to somebody; then you call them, and they tell you that the SSD part of that division is still at Samsung.

Then you give up.

[bug77]

Agree with you 100%. The twelve hard drives I have between my workstation and my home server are all spinning platter type. All are either WD Velociraptor, or WD RE4 enterprise drives. Everything is RAID-1 mirrored (using mdadm) so even if I have a failure, no activities are interrupted, no work is lost. And each of these modern platter drives is capable of > 200 MB/s sequential read, so they're certainly not slow.

The real clincher for me, is the workload rating. These WD enterprise drives are rated for 550 TBW per year workload. Most consumer SSD's are only rated for 300 TBW, 180 TBW, or even less, over their entire lifespan! So SSD's, at least the consumer variant, have a looooong way to go before catching up to spinning platters in the longevity dept.

Not only do SSD have a way to go, with each shrinking node, they actually go the other way. V-NAND mitigated this, but that's a one time "get out of jail free" card. We'll see how thing go.

Ok, better than bare controller chip at least, but without knowing the temps it reaches, we can't know if it is enough.

On desktops where the SSD is going to see more than "gaming" activity I usually recommend/install the card with a heatsink, even cheapo chinese ones are fine as it's a plain electric board and a piece of aluminum.

Anandtech didn't measure temps, but they do note:

The 2TB 960 Pro's low queue depth sequential read speed is about 300MB/s faster than the 950 Pro, once again giving Samsung the clear lead in performance and showing that the 960 Pro is significantly better than the 950 Pro where thermal limits are a factor.

in their sequential performance section.

[Delgarde]

SSDs are used in datacenters fine, not for long-term storage as HDDs are still larger and cheaper.

Cheaper, but not by as much as you'd think. Certainly the purchase price of HDD is lower - but on the scale of a datacenter, operational cost is as much of a factor, and the lower power usage of SSDs gives them a very real advantage.

As to larger, that's a balancing act... speed vs size. As you say, the trend seems to be SSD arrays for quick access to live data, slower spinning disk for long-term storage, and tape remaining desirable for backup.

[Ronshere]

I think the rapid failure of this SSD is a sort of benchmark in itself. I was going to buy one next month for a new build but now... Gotta look at other brands,

[Sonadow]

Samsung SSDs are widely used by OEMs, system builders and gamers. If it were a widespread reliability issue there would have been howls for Samsung's head already.

I'm more inclined to dismiss this as a one-off, isolated incident.

[anarki2]

I doubt you can do any of the above when the drive is not recognized in BIOS/UEFI.

You can doubt it all the way you want, that's exactly what happens.

[anarki2]

What erroneous de-initialization routines? The ones that blacklist queued trim on most samsung ssds because using it will result in data loss when it works just fine with ssds from other manufacturers? Or the speed problem problem that took samsung two tries to (badly) fix with one of their ssds a while back, with who knows what impact to longevity?

I've been burned several times before by samsung drives and their track record is not stellar, their drives may be among the fastest you can buy but I will only buy or recommend samsung drives if there is no other choice. And no, having backups doesn't count when these things always fail at the most inconvenient times.

Do bugs occur on Linux, especially with bleeding edge hardware? (yes/no)

I have no further comment.

[Mike Frett]

I've been poo-pooed here before for saying this. But I will not trust my data and/or my OS to any solid state drive for at least another 5 years of development and refinement or when even the low end drives have a minimum 5 year no questions asked warranty and replacement. Yes....yes...I know hard drives can fail suddenly way before their warranty runs out or before their rated MTBF. It has happened to me. But I am more than willing to sacrifice speed for durability and value. The cost per megabyte of SSD vs Hard Drive is still ridiculous. And hard drives have had 60 years of perfecting. Not so much SSDs.

Same here. I actually bought an ADATA that was recommended from here and it failed after after two days. I guess I've been lucky with Hard Drives, I've yet to have one fail since 1998. Others seem to have opposite experiences. USB Thumb Drives have been very good to me however =p

[niner]

Yeah. SSDs are so great for reliability and saftey of CRITICAL data. Which explains why every corporation and bank and government and university in the world uses tape and hard drives and even Blu Ray discs ( proven to be the most reliable and safe storage) for safety and reliability. Ooooooo but SSDs are soooo fast. And shiny. LOOK....squirrel!!!

Actually full flash is the hottest trend in the enterprise sector. When you got the budget, you buy flash storage. And really, when a write intensive 2 TB NVMe card costs only 3K Euros, there's not much to think about.

[droidhacker]

Agree with you 100%. The twelve hard drives I have between my workstation and my home server are all spinning platter type. All are either WD Velociraptor, or WD RE4 enterprise drives. Everything is RAID-1 mirrored (using mdadm) so even if I have a failure, no activities are interrupted, no work is lost. And each of these modern platter drives is capable of > 200 MB/s sequential read, so they're certainly not slow.

The real clincher for me, is the workload rating. These WD enterprise drives are rated for 550 TBW per year workload. Most consumer SSD's are only rated for 300 TBW, 180 TBW, or even less, over their entire lifespan! So SSD's, at least the consumer variant, have a looooong way to go before catching up to spinning platters in the longevity dept.

200 MB/s is... barely adequate. And that is only if your data happens to actually *BE* sequential. Back in the days of DOS when you could only do one thing at a time, this might have been so, but how much of the time is data ACTUALLY accessed sequentially? Get a good multi-threaded multi-processing workload going and your 200 MB/s obsolete spinny disk will turn into a head thrashing 1 kB/s.

And those workload ratings (for the obsolete spinny disk) are assuming a more-or-less SEQUENTIAL writing pattern, and do not account for MECHANICAL failures due to excessive head thrashing caused by non-sequential R/W.

In other words, when you reconsider the device's lifespan in a manner that accounts for TYPICAL usage patterns, the spinny magnet's lifespan drops like a rock, where the SSD remains solid.


And also, you're completely forgetting about the relationship between TBW and actual storage side.
TBW increases linearly with the size of the device. While your SSD that gets 300 TBW might be, say, 256 GB, that spinny magnet that gets 550 TBW is 4 TB.
300/256 = 1.17 TBW/GB.
550/4000 = 0.14 TBW/GB. Hmm, that really isn't looking so hot any more, is it?
And sure, that's 0.14 TBW/GB/year, but when the thing has a mechanical failure after 2-3 years, 5 max, then you're still only up to 0.69 TBW/GB over the device's life.