Samsung SSD 840 EVO 250GB & 1TB TLC NAND Drives Tested

MojoKid writes "Samsung has been aggressively bolstering its solid state drive line-up for the last couple of years. While some of Samsung's earlier drives may not have particularly stood-out versus the competition at the time, the company's more recent 830 series and 840 series of solid state drives have been solid, both in terms of value and overall performance. Samsung's latest consumer-class solid state drives is the just-announced 840 EVO series of products. As the name suggests, the SSD 840 EVO series of drives is an evolution of the Samsung 840 series. These drives use the latest TLC NAND Flash to come out of Samsung's fab, along with an updated controller, and also feature some interesting software called RAPID (Real-time Accelerated Processing of IO Data) that can significantly impact performance. Samsung's new SSD 840 EVO series SSDs performed well throughout a battery of benchmarks, whether using synthetic benchmarks, trace-based tests, or highly-compressible or incompressible data. At around $.76 to $.65 per GB, they're competitively priced, relatively speaking, as well."

Comments

Call me old fashion

[Taco Cowboy]

How many effective READ/WRITE cycle can the chip in SSD perform, before they start degrading ?

Has there been any comparison made in between the reliability (eg read/write cycles) of old fashion spinning-plate HD versus that of SSD ?

Re:Call me old fashion

[pipatron]

The question is still relevant. Manufacturers talk about erase cycles, but are there any massive-scale studies done by a third part on SSD failure modes?

Re:Call me old fashion

[Anonymous Coward]

Well said.

Nothing lasts forever. If a hard-driven SSD lasts 3-4 years, I don't really care that if it's used up some large fraction of it's useful lifetime, because I'm going to replace it just like I'd replace a 4 year old spinning disk.

And the replacement will be cheaper and better.

And if the SSD was used to serve mostly static data at the high speed they provide, then it's not going to have used up its write/erase cycle lifetime by then anyway.

Re:Call me old fashion

[Anonymous Coward]

Wearout is not a significant failure mode. Nearly all failures are to due to non-wearout effects such as firmware bugs and i/o circuit marginality.

Re:Call me old fashion

[beelsebob]

Yes, many sites have done the maths on such things. The conclusion "finite life" is not the same thing as "short life". SSDs will in general, outlast HDDs, and will in general die of controller failure (something which affects HDDs too), not flash lifespan.

The numbers for the 840 (which uses the same flash, with the same life span) showed that for the 120GB drive, writing 10GB per day, you would take nearly 12 years to cause the flash to fail. For the 240/480/960 options for the new version you're looking at roughly 23, 47 and 94 years respectively. Given that the average HDD dies after only 4 years (yes yes yes, we all know you have a 20 year old disk that still works, that's a nice anecdote), that's rather bloody good.

Re:Call me old fashion

[beelsebob]

The problem with large scale studies on this is that it takes too long to happen to actually study. You need to study real world usage patterns, and in the real world it takes decades before the flash actually wears to death. Controller failure (as is possible with HDDs too) will generally happen long before the flash becomes unwritable.

Re:Call me old fashion

[beelsebob]

Yes, they were solved in a firmware patch a long time ago.

Re:Call me old fashion

[beelsebob]

The problem with such tests of writing as much as you can as fast as you can is that they're rather deceptive. They don't allow TRIM and wear levelling to do their thing (as they normally would), and hence show a much worse scenario than you would normally be dealing with. Actual projections of real life usage patterns writing ~10GB to these drives per day show you can get their life span in years (specifically the 840 we're talking about here) by dividing the capacity (in gigabytes) by 10.

Re:Call me old fashion

[Joce640k]

Modern SSDs have big capacitors in them to avoid that (well, some of them do...). They can complete pending writes on capacitor power alone.

Re:Call me old fashion

[Gaygirlie]

How many effective READ/WRITE cycle can the chip in SSD perform, before they start degrading ?

They don't start degrading, per se. Performance-degradation is all due to wear-levelling and the amount of free blocks on the drive, and that varies between manufacturers. Generally the advice is to have atleast 20% of the drive free at all times for wear-levelling and TRIM to work efficiently and in such a situation there should be no performance-degradation.

As for reading and writing cells? Well, you can read a cell indefinitely. You cannot write to cells forever, however, and once the limit comes there is 100% degradation -- so to speak -- as in that that cell cannot be written to ever again. It just goes from 100% to 0%, so using the term "degradation" for that still seems useless. I'll repeat, though, that it can still be read from even if it can't be written to.

Has there been any comparison made in between the reliability (eg read/write cycles) of old fashion spinning-plate HD versus that of SSD ?

Plenty, but how much those comparisons actually cover and how reliable they are is subject to debate. Generally the consensus is that SSDs are more reliable nowadays as full-on controller-failures are very rare and since the SSDs can still be read from even if they hit the maximum amount of writes that means your data is quite a lot safer in the long run -- if a regular, mechanical drive can't write to some sector it most likely can't read it either, and that means your data is as good as gone.

Re:Call me old fashion

[AmiMoJo]

It depends on what you use it for. I managed to wear out an Intel XM-25 160GB SSD a few years ago by hitting the 14TB re-write limit.

Modern SSDs so a lot of compression and de-duplication to reduce the amount of data they write. If your data doesn't compress or de-duplicate well (e.g. video, images) the drive will wear out a lot faster. I think what did it for me was building large databases of map tiles stored in PNG format. Intel provide a handy utility that tells you how much data has been written to your drive and mine reached the limit in about 18 months so had to be replaced under warranty.

Re:Call me old fashion

[dmesg0]

TLC endurance was tested in this article:

http://uk.hardware.info/reviews/4178/10/hardwareinfo-tests-lifespan-of-samsung-ssd-840-250gb-tlc-ssd-updated-with-final-conclusion-final-update-20-6-2013

Re:Call me old fashion

[Tapewolf]

Ok, you're old fashioned.

This was a thing, yes, but only for that brief period when you actually got your slashdot id. Since then? Not so much ...

--Q

Technically it becomes less and less reliable each time they do a die shrink on the flash. Adding a whole extra bit level makes things worse still. In the early 2000s you were looking at 100'000 P/E cycles, maybe a million for the really good stuff. Good TLC memory seems to be rated around 3000, with a figure of 1000 being widely quoted, and in some cases, less.

Realistically, they've designed the drive to fight tooth and nail to avoid doing rewrites, and in actual fact it looks like they've put a layer of fast SLC cache in front (i.e. the million-cycle stuff). What could be more interesting is the retention period - if the thing is left powered off for three months it could well be left unreadable.

Re:Call me old fashion

[Rockoon]

Intel provide a handy utility that tells you how much data has been written to your drive and mine reached the limit in about 18 months so had to be replaced under warranty.

You were (amplified?) writing 32.8 GB per day, on average.

Clearly you will run into SSD erase-limit problems at such a rate, but such workloads normally turn out to not be tasks that actually benefit from an SSD to begin with (32.8GB/day = 380KB/sec, so the devices speed wasnt actually an issue for you)

You were either very clever and knew you would hit the limit and get a free replacement, or very foolish and squandered the lifetime of an expensive device when a cheap deice would have worked.

In any event, in general the larger the SSD the longer its erase-cycle lifetime will be. For a particular flash process its a completely linear 1:1 relationship, where twice the size buys twice as many block erases (a 320GB SSD on the same process would have lasted twice as long as your 160GB SSD with that work load)

Re:Call me old fashion

[Anonymous Coward]

Power failure?

You don't have a UPS or other standby power source available? You know its 2013 right...

Willing to spend hundreds on an ultra fast STORAGE device and have no backup power available? really? come on...

That's some messed up priorities there... Spend a hundred bucks on a UPS already.

Then you don't ever have to worry about data corruption. Or the much more common... Loss of unsaved work due to power failure...

Re:Call me old fashion

[Rockoon]

Indeed. I just don't see how the erase-limit issue applies for most people. The most common activity where it might apply is in a machine used as a DVR (dont use an SSD in a DVR), with the next being a heavily updated database server (you may still prefer the SSD if transaction latency is important.)

For people that use their computers for regular stuff like browsing the web, streaming video off the web, playing video games, and software development.. then get the damn SSD -- its a no-brainer for you folks.. you will love it and it will certainly die of something other than the erase-limit long before you approach that limit.

Re:Call me old fashion

[Rockoon]

Technically it becomes less and less reliable each time they do a die shrink on the flash. Adding a whole extra bit level makes things worse still. In the early 2000s you were looking at 100'000 P/E cycles, maybe a million for the really good stuff. Good TLC memory seems to be rated around 3000, with a figure of 1000 being widely quoted, and in some cases, less.

Lets not neglect the fact that while every die shrink does reduce the erase-limit per cell, it also (approximately) linearly increases the number of cells for a given chip area. In other words, for a given die area the erase limit (as measured in bytes, blocks, or cells) doesnt actually change with improving density. What does change is overall storage capacities and price.

When MLC SSD's dropped from ~2000 cycles per cell to ~1000 cycles per cell, their capacities doubled (so erases per device remains about constant) and prices also dropped from ~$3/GB to about ~$1/GB. Now MLC SSD's are around ~600 cycles per cell, their capacities are larger still (again erases per device remain about constant), and they are selling for ~$0.75/GB (and falling.)

By every meaningful measure these die shrinks improve the technology.

So now lets take it to the (extreme) logical conclusion, where MLC cells have exactly 1 erase cycle (we have a name for this kind of device.. WORM: Write Once Read Many.) To compensate, the device capacities would be about 600 times that of todays current capacities, so in the same size package as todays 256 GB SSD's we would be able to fit a 153 TB SSD WORM drive, and it would cost about $200.

Re:Call me old fashion

[hackertourist]

(32.8GB/day = 380KB/sec, so the devices speed wasnt actually an issue for you)

That's an odd way to look at it. You assume that GP spreads out his writes evenly over 24h, and has no desire to speed things up.

Re:Call me old fashion

[AmiMoJo]

In my case having an SSD made a huge impact. I was using offline maps of a wide area build from PNG tiles in an sqlite database with RMaps on Android. Compiling the databases was much faster with an SSD. I was doing it interactively, so performance mattered.

I can only tell you what I experienced. I installed the drive and I didn't think about it wearing out, just carried on as normal. The Intel tool said that it had written 14TB of data and sure enough writes were failing to the point where it corrupted the OS and I had to re-install.

I was using Windows 7 x64, done as a fresh install on the drive when I built that PC. I made sure defragmentation was disabled.

I'm now wondering if the Intel tool doesn't count bytes written but instead is some kind of estimate based on the amount of available write capacity left on the drive. I wasn't monitoring it constantly either so perhaps it just jumped up to 14TB when it noticed that writes were failing and free space had dropped to zero.

It was a non-scientific test, YMMV etc etc.

Re:Call me old fashion

[Rockoon]

So what exactly is the difference between a write cycle and an erase cycle in practical terms?

The difference is that there is no such thing as a "write cycle." The guy that you are replying to doesnt actually know much about what he is talking about.

In regard to the general difference between writes and erases the terminology on the table are supposed to be write amplification, block size (typically 256KB), and page size (typically 4KB.) Write amplification occurs when data smaller than a page is frequently written or "modified."

In practice write amplification is typically below 2x on modern controllers, and obviously always larger than 1x.

Worst case write amplification is horrendous at 4096x but the typical scenario where near-worst-case amplification does occur turns out to be low volume traffic in practice (for every near-worst-case page write the SSD experiences, it experiences thousands of near-best-case page writes.)

The worst cases turn out to be log file writes, but thats very small change compared to other write activity (its like worrying about that $1 surcharge on your $150/month cable bill.)

Re: Call me old fashion

[Zeinfeld]

Hmmm I replace my hard drives when I start to see RAID errors. I don't plan to run SSD raid as the on board fault tolerance should be ok.

Would be nice to have hard data on expected failures so that I know whether to plan for a three or a six year lifespan. I generally replace my main machine on a six year cycle as I have a lot of expensive software. Looking to upgrade this year when the higher performance intel chips launch.

1tb is quite a lot. Probably more than I need in solid state. The price is also quite a bit more than the $0.05/gig for Hard drives. But it's getting a lot narrower. And RAID 1 doubles that cost anyway...

Re:Call me old fashion

[TheLink]

Just don't buy OCZ crap?
http://www.behardware.com/articles/881-7/components-returns-rates-7.html
http://www.behardware.com/articles/881-6/components-returns-rates-7.html

It's not a study but it's good enough for me.

Re:Call me old fashion

[0123456]

It would be better if, like hard drives, SSDs were simply designed not to die in the event of unexpected power failure.

About 80% of the hard drive failures on our servers over the last few years have been due to power failures. They run fine for years, then the power goes out and they're dead on boot.

So 15k HDDs don't seem to like power failures either.

Re:Still put off by price. :(

[Gaygirlie]

Note: Anyone here about any programs like spinrite or other for drive recovery for SSD's?

There is no such a thing except for the few that are just trying to dupe you into giving them money. Why? Well, as long as the drive's controller itself is working and the drive's internal state isn't corrupted you can read the cells indefinitely. You cannot write to cells indefinitely, but all major manufacturers these days promise that even if all the cells failed in the whole drive you should be able to read them. On the other hand, if the drive's controller goes bonkers or the internal state gets messed up there is *no software whatsoever* that can fix it. You'd have to open the drive and work with the actual flash-chips themselves in the hopes of recovering your data, and due to the nature of SSDs where the cells can be re-located at any given time for wear-levelling purposes that'd be one helluva task.

Now, if your filesystems or such get messed up any tool that works on mechanical HDDs works just fine on SSDs. There is no difference.

Re:Not paying for TLC at that price

[beelsebob]

Except if you actually bothered to educate yourself, you'd find that at the capacities samsung is offering you, if you write to them at 10GB a day, every day, they'll last entirely respectable times (12,23,47,94 years respectively for 120,240,480 and 960GB drives).

Hot vs Crazy

[bdwoolman]

Here's the thing. SSDs are now more reliable than when this guy logged this report.

But are still maybe not as steady Eddie as a good-quality HDD. But we still want them because having an SSD boot drive changes the whole computing experience due to their awesome speed. And since we are good about backups (Are we not?) we can be relaxed as we ride the SSD smokin' fast Roller Coaster. SSD or HDD then what's the problem if we have data security. Both are gonna FAIL. So what if Miss SSD stabs me for no good reason? It was a helluva ride, Bro. And well worth the stitches. I do wish SLC NAND was not priced out of reach, but, hey, when it comes to hottness we take what we can get. Right?

Okay. This is Slashdot we get no hottiness...no hottiness at all.. No no no hottiness. It's pathetic really. ....

Only some do

[Sycraft-fu]

New Intel drives do, as they use the Sandforce chipset. However Samsung drives don't. Samsung makes their own controller, and they don't mess with compression. All writes are equal.

Also 14TB sounds a little low for a write limit. MLC drives, as the XM25 was, are generally spec'd at 3000-5000 P/E cycles. Actually should be higher since that is the spec for 20nm class flash and the XM25 was 50nm flash. Even assuming 1000, and assuming a write amplification factor of 3 (it usually won't be near that high) you are talking 52TB if the drive has no internal overprovisioning, which it probably does.

As an example, AnandTech tested a Samsung 840 TLC drive. The 250GB drive was able to take about 266TB of incompressible data, which translates to a bit more than 1000 P/E cycles.

If you have a high write workload, their MLC drives aren't that much. A 512GB 840 Pro drive will run you like $450. That should get you somewhere in the realm of 1.5PB of writes before it fails, maybe more.