When SSD Performance Goes Awry: Samsung's SSD 840 read performance degradation explained

[Per Hansson]

An avalanche of reports started to surface last September when users began to notice that their usually speedy Samsung SSD 840 and SSD 840 EVO drives were no longer performing as they used to.

The issue has to do with older blocks of data: reading old files would be consistently slower than normal – as slow as ~30Mb/s – whereas any newly written files, like the ones used in benchmarks would perform as fast as new – around 500Mb/s for the highly regarded SSD 840 EVO. The reason no one had noticed before (we reviewed the drive back in mid-September 2013) is that data has to be several weeks old to show the problem. Samsung promptly admitted the issue and proposed a fix.

We've looked deep into the problem to understand what's been fixed and what hasn't. Samsung owes its customers an explanation.

Read the complete article.

 

[wujj123456]

Finally a hero! An editor actually cares about the unfortunate minority that bought original 840 early!!!

I complained about my vanilla 840 as well, but hell no other site even mentioned the problem is universal to TLC, not just EVO. Most of original 840 users' voices are swamped by EVO users. Then Samsung just shamelessly denied the problem.

Anyway, my latest SSDs are all Sandisk's MLC (Extreme II and Extreme Pro), and I am very happy about them. They seem to be performing better than even my 840 Pro after using for quite a while.

 

[Per Hansson]

As a side note: It should be noted that to get the best possible performance the 960GB 845DC EVO uses 8x128GiB K90KGY8S7M-CCK0 NAND chips. The 480GB model uses 4x of the same NAND chips. But having just 2 chips on the 240GB model would have resulted in poor performance, so it uses 4x64GiB K9CHGY8S5M-CCK0 NAND chips instead.

The configuration in the 845DC EVO’s OEM variant PM853T is exactly the same. The 840 EVO also uses 64GiB K9CHGY8S5M-CCK0 NAND in its smallest 120GB variant. The 250GB, 500GB & 1TB drives use K90KGY8S7M-CCK0 NAND. The 750GB model has a mix of 4xK90KGY8S7M-CCK0 128GiB and 4xK9CHGY8S5M-CCK0 64GiB NAND due to its odd size.

 

[DaveBG]

So what is the conclusion ? Get Intel 750 4D nand and forget about it.

 

[Guest]

I have a samsung 840 evo 120gb. Crazy thing is that I have 496MB/s write speed(max is 410 lol) and read speed worse at 417MB/s.

 

[abysal]

I think it's time to boycott Samsung for this user. I don't appreciate companies lying, Samsung should of owned up to the issue sooner. And they're still denying some of the affected TLC NAND product line it seems.

 

[Antykain]

Just ran the Samsung Magician Benchmark test on my 120GB 840 EVO, and it seem to be running normally.. 549/530. A tab faster than my pair of 500GB 850 EVO's which ran 535/529. Latest firmware on both.

 

[Evernessince]

Just ran the Samsung Magician Benchmark test on my 120GB 840 EVO, and it seem to be running normally.. 549/530. A tab faster than my pair of 500GB 850 EVO's which ran 535/529. Latest firmware on both.

The samsung magician benchmark is garbage compared to other benchmarks. It said my 850 was getting 30k iops while I was getting 90 - 110k in Crystal disk.

 

[Steve]

Just ran the Samsung Magician Benchmark test on my 120GB 840 EVO, and it seem to be running normally.. 549/530. A tab faster than my pair of 500GB 850 EVO's which ran 535/529. Latest firmware on both.

I have the 120GB and 240GB Evo models in two separate HTPC's. Both systems work really well, they are only Core i3 powered. I will have to check the performance but without benchmarking I haven't noticed a difference. It will be interesting to check them out when I get a chance.

 

[Guest]

I've been avoiding Samsung drives on my home PCs because I noticed that all the 840 SSD drives my work bought and installed on the work PCs were pretty slow (over 200 PCs). Sometimes, they would just stall for a couple of seconds while loading files.

I kept telling people that "something was up. SSD drives should not behave like this" but I guess it wasn't annoying enough for people to do something about it.

Anyhoot! I use Crucial SSDs (I forget the actual model #) for my home PCs and they've worked flawlessly from day one.

 

[Guest]

I am planning to buy a Samsung 850 model of some description shortly. I sent a query to Samsung to see if they were affected. Their reply is:

"Actually the drop down in performances was related to the 840 EVO, but it has just released a new firmware version which fixed the problem. The issue is not related to any other model, so the 850 (EVO and PRO) are not affected."

Cheers,
Don

 

[Guest]

Note: Normal benchmark programs will NOT show the problem. That's because they always write new files. And new files are not affected. But the older the files the slower they will be read. Because the voltages in the TLC cells change over time and the stored values are more difficult to identify.

To test for the problem, use 'ReadSpeedTester' or 'HD Tune' linked in the article.

 

[Divine_Assassin]

As the owner of a (vanilla) Samsung 840 120GB this sort of crap is the reason why I went with a Crucial MX100 512GB when it came time upgrade my rig.

 

[melkiik]

Way to go Hansson, nice article! And the graphics too, awesome.

I myself do own 2 Samsung 850's and they work like a charm. Pity you guys who wen't for the early birds :c

 

[Sule Song]

Hi... My computer just crashed, my C drive Samsung SSD 840 PRO 512GB is not showing any life. My D and E drives, 2TB each are now working. This aging problem, could that be the reason.

 

[Lionvibez]

So what is the conclusion ? Get Intel 750 4D nand and forget about it.

No the conclusion is if you must go with a samsung drive make sure its SLC or MLC and stay away from their TLC drives until they have permanently fixed the issue.

Stuff like this is why my rig only has Intel SSD's currently and when its time to upgrade it will be intel as my first choice and maybe a 850pro I won't touch anything with TLC anyways.

 

[Per Hansson]

Just ran the Samsung Magician Benchmark test on my 120GB 840 EVO, and it seem to be running normally.. 549/530. A tab faster than my pair of 500GB 850 EVO's which ran 535/529. Latest firmware on both.

That benchmark, or any like it which creates a new test file will obviously not show the issue since it only occurs to old data that has been on the drive for some time, as I explained in the article.

So what is the conclusion ? Get Intel 750 4D nand and forget about it.

That drive has terribly poor low I/O transfer size and low queue depth transfer size performance so for a consumer workload I would not recommend it.

 

[bazz2004]

There was supposed to be a new fix out for the Samsung 840 Evo drives in March but it hasn't surfaced. Everything has gone very quiet so maybe they can't come up with a 100% solution.

 

[aki009]

Cool article.

FWIW, some "older" files are still fast(er) likely due to the drive rewriting blocks as part of erase block management and wear leveling.

 

[thorpj]

Just ran the Samsung Magician Benchmark test on my 120GB 840 EVO, and it seem to be running normally.. 549/530. A tab faster than my pair of 500GB 850 EVO's which ran 535/529. Latest firmware on both.

So what is the conclusion ? Get Intel 750 4D nand and forget about it.

No the conclusion is if you must go with a samsung drive make sure its SLC or MLC and stay away from their TLC drives until they have permanently fixed the issue.

Stuff like this is why my rig only has Intel SSD's currently and when its time to upgrade it will be intel as my first choice and maybe a 850pro I won't touch anything with TLC anyways.

Will the 850 evo have the same problem? I'm thinking of getting a refund for my 840 evo...

 

[Darth Shiv]

Realistically they should trade in 850 EVOs for 840 EVOs for free. If they are confident the 850s don't suffer the same issue that is...

 

[Guest]

See my earlier post.
Don

 

[Tgard]

So what is to be our final solution for this? Samsung will keep making new firmware that wears out our drives faster than normal until the warranty has expired? The really do need to replace our 840 Evos with 850s if they truly do not have the problem. Class action for this maybe? It would be a lose-lose either way as it would probably take years and then all owners that have their names in will get a check or discount for a new drive in the whopping amount of $4.78 for every drive you own, 6 years after the fact.

To combat the problem, I have been making a ghost image of the drive and then formatting and writing the whole image new every couple months. (Well twice so far). Not really interest in a firmware fix that keeps re-writing the data over and over and lessening the re-writes on the drive.

My first SSD was a Samsung Evo, and since then I have gone with Intel and Crucial. Samsung has probably lost a customer in me. They need to stop playing around and make this right. If we would boycott, then we may get some satisfaction.

 

[DAOWAce]

Well, now I wish I picked up an 850 EVO instead of a second, cheaper 840 EVO.

Respect lost for Samsung; but honestly people have been wary of TLC NAND since its debut, now it's completely proven it's not the way to go for anything but budget drives.

When finances become better, I'm definitely replacing my 840 EVOs (and Vertex3/4) with something with a proven reliability record, but until then, at least nothing's close to failing yet and performance has yet to drop since the first firmware update, but I'll be watching..

 

[SuperVeloce]

So what is to be our final solution for this? Samsung will keep making new firmware that wears out our drives faster than normal until the warranty has expired? The really do need to replace our 840 Evos with 850s if they truly do not have the problem. Class action for this maybe? It would be a lose-lose either way as it would probably take years and then all owners that have their names in will get a check or discount for a new drive in the whopping amount of $4.78 for every drive you own, 6 years after the fact.

To combat the problem, I have been making a ghost image of the drive and then formatting and writing the whole image new every couple months. (Well twice so far). Not really interest in a firmware fix that keeps re-writing the data over and over and lessening the re-writes on the drive.

My first SSD was a Samsung Evo, and since then I have gone with Intel and Crucial. Samsung has probably lost a customer in me. They need to stop playing around and make this right. If we would boycott, then we may get some satisfaction.

The "problem" here is that the consumer ssds (840evo) will survive somewhere around 10x the number of writes Samsung states for warranty. It's more likely something other than flash will die first (controller?). You would need a high server load to wear those drives in 3 years time anyway.