TR's SSD endurance test

[P5-133XL]

I thought people might be interested:

Tech Reports SSD endurance test - Intro
22TB update
200TB update
300TB update
500TB update
600TB update
1PB update: 3 drives dead, 3 still alive

 

[LunarMist]

So the 1TB drives should be worth 800TB of writes at least.

 

[CougTek]

I've been telling people to buy the Corsair Neutron GTX since it launched. I'm very glad to see those test confirm what I strongly suspected. It has one of the best garbage collection implementation out there and it pays. I've even used it in a few servers and so far, no issue.

 

[jtr1962]

My 240GB Intel 330 has only 2.05 TB of writes of it after a bit over a year in service. It's looking like it'll last at least 100 years at that rate. I wonder if flash memory cells also degrade with time?

 

[LunarMist]

My 240GB Intel 330 has only 2.05 TB of writes of it after a bit over a year in service. It's looking like it'll last at least 100 years at that rate. I wonder if flash memory cells also degrade with time?

I wrote 2TB on a single SSD in one day.

 

[LiamC]

What do you do with them?

 

[Mercutio]

What do you do with them?

That's not terribly difficult. A cache drive in a ZFS storage pool of sufficient size; a drive that's acting as a target for disk images; somebody who dumps a vacation worth of .RAW files.

I've been putting 250GB SSDs in servers as a sort of intermediate upgrade to a full hardware replacement. So far, that's a strategy that's worked beautifully for me. I've not had an issue on database/web servers to date, but I have had a couple end users kill more than one SSD what I think is an unreasonably short time frame. I don't know if they've just had bad luck or if they do things that kill drives. I definitely know users who sit and watch Windows run a defrag every day now that I think about it...

 

[Stereodude]

I wonder if flash memory cells also degrade with time?

The data in them sure does.

 

[Bozo]

Early in the SSD testing write up, they mentioned that flash memory exhibits the same degradation.

 

[LunarMist]

What do you do with them?

I download files. Isn't that what the drives are for?

 

[LunarMist]

I've been telling people to buy the Corsair Neutron GTX since it launched. I'm very glad to see those test confirm what I strongly suspected. It has one of the best garbage collection implementation out there and it pays. I've even used it in a few servers and so far, no issue.

I wonder if they can get 1TB from the LAPD controller?

 

[CougTek]

I wonder if they can get 1TB from the LAPD controller?

I didn't know that the dispatcher from the Los Angeles Police Department was selling drive.

Regarding the LAMB controller, it would probably be possible, but I'm not aware of any such drive.

 

[LunarMist]

:lol: Yeah, that's what I mean.

 

[LunarMist]

So why is the Samsung 840 Pro weaker than others? I know the regular 840 used TLC memory but the Pro doesn't.

 

[Stereodude]

So why is the Samsung 840 Pro weaker than others? I know the regular 840 used TLC memory but the Pro doesn't.

The test is largely pointless. Those parameters don't mean anything. There is nothing that suggests a 50 "score" on a Samsung = a "50" score on a Corsair.

 

[LunarMist]

The test is largely pointless. Those parameters don't mean anything. There is nothing that suggests a 50 "score" on a Samsung = a "50" score on a Corsair.

It seems to me that they should test the drives to the point of data errors/other failure.

 

[Mercutio]

It seems to me that they should test the drives to the point of data errors/other failure.

Does that prove anything for any more than the drives in the sample?

 

[Chewy509]

Does that prove anything for any more than the drives in the sample?

My thoughts as well. To make that meaningful you would need at least 1000 units of each drive, and test them all to death before you could draw any correlation or conclusions.

 

[P5-133XL]

They are testing to failure: The last article is just update at the 200TB point. There will be more updates, so stay tune...

I agree that statistically, the tests are likely worthless because a sample size of one each is far too small. That being said, even anecdotal evidence is information that most SSD users would like to know and can't conveniently get before buying.

 

[LunarMist]

Yeah, they should start with 13 and go from there.

 

[Stereodude]

I don't mean to further muddy the waters, but their test is invalid anyhow. The faster you write data to the drive the less resting time the flash cells have between program and erase cycles. This will exaggerate the wear on the flash cells more than normal use would.

Doing this right would involve creating Vtm/V curves for the flash cells and characterizing the voltage levels / gate charge on the flash cells. Of course that's way beyond the scope of what a review website could do.

 

[CougTek]

I don't mean to further muddy the waters, but their test is invalid anyhow. The faster you write data to the drive the less resting time the flash cells have between program and erase cycles. This will exaggerate the wear on the flash cells more than normal use would.

This is only true for a user environment. If you put the drive in a busy database server, this will be its usage pattern. Granted, those aren't SAS drives and shouldn't be use in such an environment, but I'm sure they'll be anyway since several companies don't have a lot of monetary ressources but still want or need the increased performances SSDs provide. I'm almost certain that many of the Asus/Gigabyte/MSI/Tyan/etc servers out there, which don't require rebranded hard drives from the same company in order to respect the warranty (like HP and IBM do, for instance) will have consumer drives into them.

 

[Stereodude]

This is only true for a user environment. If you put the drive in a busy database server, this will be its usage pattern.

I wouldn't expect a busy database server to be continuously writing data to the drive as fast as possible, erasing it and starting over again 24/7 either.

 

[Mercutio]

No, but that's what a dedicated drive for transaction logs might be doing.

 

[timwhit]

No, but that's what a dedicated drive for transaction logs might be doing.

Or app/web server logs.

 

[P5-133XL]

300TB update

 

[LunarMist]

300TB update

So it looks like the TLC is good for at least 1200 write cycles (300TB/250GB). Even if it only safe to assume 1/3 of that (5400 cycles) a 1TB drive should be good for 400TB.
I'll never understand why the Micron 960GB is only rated 72TB, but I will probably buy more of those before an EVO.

 

[Handruin]

What's wrong with the EVO?

 

[Stereodude]

What's wrong with the EVO?

It uses TLC flash.

 

[LunarMist]

The problem is that the TLC drives should be much cheaper if only half the amount is needed.

 

[Handruin]

It uses TLC flash.

It functions OK and performs well. Is it still bad because it has TLC flash?

 

[Handruin]

The problem is that the TLC drives should be much cheaper if only half the amount is needed.

All flash drives should be cheaper. They've been out long enough now.

 

[Stereodude]

It functions OK and performs well. Is it still bad because it has TLC flash?

It definitely has less endurance and certainly will have worse retention as well.

 

[Stereodude]

Perhaps I should clarify. My beef is flash with 3 bits per cell flash. Trying to keep 8 distinct voltage levels in a flash cell is a disaster waiting to happen.

 

[Handruin]

Perhaps I should clarify. My beef is flash with 3 bits per cell flash. Trying to keep 8 distinct voltage levels in a flash cell is a disaster waiting to happen.

Have there been problems documented as a result of this? I get that on paper they appear less durable and SLC and MLC are more desirable but I haven't seen large or many reports of the EVO with TLC being hugely problematic.

 

[Stereodude]

Have there been problems documented as a result of this? I get that on paper they appear less durable and SLC and MLC are more desirable but I haven't seen large or many reports of the EVO with TLC being hugely problematic.

Well, the SSD with 3-bit per cell flash is the first one to have a meltdown in the TR SSD endurance test. Average users probably haven't gotten to that point yet.

 

[LunarMist]

No and they probably won't into endurance problems very often. The heavy users would be more likely to buy other brands or the MLC pro series for performance reasons as well.

 

[Handruin]

No and they probably won't into endurance problems very often. The heavy users would be more likely to buy other brands or the MLC pro series for performance reasons as well.

I got mine Sept 13th of this year and so far I've written 1.24TB to the drive. I'll be curious to see how well it holds up. I don't think I'm a heavy user of the drive at these rates. If I were a heavy user I would consider an MLC or SLC (if they can be found anymore in reasonable sizes).

 

[P5-133XL]

500TB update

 

[time]

That's really, really impressive. I'm starting to think it will take a controller failure to take out one of these current SSDs.

Now that more cycles have been completed, I'm not worried about the Samsung 840 like the author is. It's using up its overprovisioning at a slow - and more importantly steady - rate, but after 500TB the remaps only add up to 1% of the nominal capacity. Given that overprovisioning is probably 20% on this TLC drive, it may well make it to 1 petabyte, at which point I don't think anyone cares.

I notice that the Corsair Force boot drive did its well-known suicide bomber routine - surprised they seemed unaware of its reputation. Hopefully that didn't contribute towards the temporary aberration on the Samsung drive after restarting the system to test data retention?