Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
Samsung 840, the first TLC SSD, arrives
My view for a while now has been that SSD performance is good enough. We have good controllers from Marvell, SandForce, Samsung, and now LAMD. Indilinx's upcoming Barefoot 3 controller will probably join them, and there's also the possibility that more will come out of nowhere like the LAMD controller recently did. There are still some bad controllers on the market, but there are enough good ones and enough competition at all stages that no one can gouge you with monopolistic prices.
Sure, you can run benchmarks and find that this SSD is 20% faster than that one in such and such particular scenario. But that doesn't matter anymore, at least for consumer use, as the good ones are all so fast. Debating whether you need this SSD or that one for performance reasons is like asking whether a GeForce GTX 680 or a Radeon HD 7970 GHz Edition is better for watching videos on YouTube.
Reliability still matters, but there are plenty of options that are both reliable and fast. That leaves the price tag as the only good reason to get a hard drive anymore. If you need a terabyte or more of capacity, hard drives are cheaper by more than an order of magnitude.
Initially, there was SLC NAND flash. This is NAND flash in which each cell can have two different levels of charges stored, so that the cell can hold one bit of information. NAND flash moved to MLC a long time ago, and SSDs quickly adopted it. This means cells have possible levels of charge, so that each cell can hold two bits of information. Today, there are only a handful of SLC SSDs.
TLC means that a cell can have eight possible levels of charge, which allows a single cell to hold 3 bits of information. Having 3 bits in the same space that used to have 2 bits means you an give 50% more capacity for a given cost of production. That's not going to close the price gap with hard drives all by itself, but it's movement in the right direction.
While TLC NAND flash has been available for quite a while, there are problems with reliability. Eight possible levels of charge means narrower gaps between each possible amount of charge. The actual charge on a cell is some integer multiple of the charge of an electron, of course, but the way NAND flash works is to measure the charge and then round it to a charge level that corresponds to a particular combination of bits.
The problem is that you can't have an exact charge level stay that way forever. For reasons of quantum mechanics, electrons leak out (or in) and change the charge level of a cell. If this happens to too many electrons in the same cell, reading the data later will give the wrong value. The closer the meaningful levels of charge are to each other, the fewer electrons have to leak out to give wrong data.
There are a variety of ways to combat this to some degree. There is a lot of redundancy built into NAND flash, just like hard drives, so even if one bit does get flipped, it will be detected and corrected. The state of various cells is periodically refreshed, so that even if the charge has drifted slightly from the intended value, it will be erased and rewritten with the proper charge, and usually before it can cause the wrong data to be read. That's not new to TLC NAND flash, as SLC and MLC had to do so as well.
But TLC requires more of it to maintain reliability. And SSDs are generally believed to require much greater reliability than a simple USB flash drive that you buy for $10. Losing your OS, programs, and primary copy of your data is a much bigger problem than having a USB flash drive die, since you probably weren't relying on the latter for anything critically important.
Samsung isn't the only SSD vendor that will use TLC NAND flash, of course. The cost advantage is large enough that they'll probably all jump on board soon. OCZ has been talking about it for a while already, though they commonly talk about upcoming products months before they launch. But today's public launch of the first TLC SSDs, and from a reputable vendor at that, is a major sign of progress.
As with die shrinks, Samsung says not to expect prices to crater instantly. It takes a while to ramp up production of new products. But this should eventually mean that you get 50% more SSD capacity for a given price tag than you would have with MLC flash. And that's a big deal.
Comments
They've done it for a long time with USB flash drives. The NAND flash chips are designed to be TLC, so if you want to make, say, a 32 Gb chip, you don't need exactly 11453246123 2/3 cells. You go over that slightly and waste a handful of extra bits, and then sell it as a 32 Gb chip.
NAND flash doesn't need the speedy random access of DRAM. I'd assume that DRAM chips are always powers of 2 of everything because there are low level optimizations that make random reading and/or writing faster, though I don't know that for certain. If you can access memory 1 ns faster in DRAM, that's a big deal. 10 ns could be the difference between blazing fast and unsellable garbage. But for NAND flash, a 10 ns difference in random access time is basically a rounding error.
It may be harder for most people to conceptualize, but electronics don't care. If you think about it a bit differently:
Instead of looking at it bit by bit, with each storage device holding exactly one bit, you just conglomerate them. Each storage "unit" holds 24 bits (3 bytes), and consists of a group of 8 tri-state diodes. The drive firmware usually doesn't get asked for a single bits worth of information at a time anyway - usually it's in chunks of 512B-32k (file system cluster sizes), so it just fills up it's buffer with 3B chunks at the lowest level, rather than single bit chunks. It probably does this at some much larger size than 3B anyway.
You do realize that the only benefit of TLC is cost cutting right? SLC is still the fastest type of NAND FLASH memory, it is also the most enduring one.
TLC offers about 40% of the performance you would get with SLC, and much lower endurance(less IOP Life Time, especially for tunnel-injection which is used for write operations) which means you DO NOT want it in your HDD, it will mainly be used for thumb drives, and other non critical applications, and for low-end and cheap HDD's that would be sold to the uneducated masses.
I prefer to think of it as more capacity for the same price rather than a lower price for the same capacity. But yeah, that's the point of TLC.
And it's not just that it "will be" used in USB flash drives and such. It already is. They already have four bit per cell NAND flash, too. What's new here is TLC in a real SSD.
Cost cutting is a big deal, too. The only reason people still use hard drives is the price tag. (Well, I guess I shouldn't discount ignorance.) So long as SSDs continue to scale with Moore's Law, more and more people will decide to buy just an SSD and skip the hard drive. A few years from now, if you can get a ~500 GB SSD for $100, do you really need a hard drive?
There's also the question of, suppose that you're only willing to spend $80 for an SSD, and want to know what you can get with it. It used to be that you couldn't get anything decent for that price. Then you could get ~30 GB and install your OS. Then you could get ~60 GB and install a bunch of programs in addition to your OS, though generally not large once. Now you can get ~120 GB and install nearly whatever you want, though not necessarily everything you'd like to install all at once. Give it a couple more years and you'll be able to get ~240 GB and install everything on the SSD without having to prioritize much. The more you can fit, the more tempting it becomes.
Hello Quizzical, always enjoy reading your posts. Been so long since I've posted (if I ever have) that I had to retrieve my password so that I could ask you a question. I build my own computers, and on my current I used a crucial 256 ssd, a sata 3 one (they had just come out), as my boot disc, and keep games and music on a 1T sata 3 hard drive.
I have heard of some new type of use for an ssd...something to do with holding oft-used info or something to speed things up? Perhaps as a type of virtual ram or something? Do you know anything about this? Wondering if its something that would offer me more speed. Not sure if it's a regular ssd or some new type with a defined purpose.
Thanks for any answer!
jeff in san diego
edit: ah i see it is my fifth post
You're probably thinking of using SSDs as cache rather than as a normal drive.
There are two basic things that you can do with an SSD. One is that you can use it about like you would use a hard drive, except that it will be a lot faster. Naturally, you prioritize some by putting programs on the SSD and videos or music or other bulk data on your hard drive, but you still use the SSD fundamentally like you would use a hard drive--as opposed to system memory, a tape backup drive, a DVD drive, etc.
The other is that you can use the SSD as cache for your hard drive. The idea is that software watches which hard drive sectors get loaded from your hard drive a lot and stores them on the SSD. Then in the future, when it wants to load something, it checks first to see if it's on the SSD. If it is, it grabs it there and doesn't touch the hard drive. If not, then it has to load the data from the hard drive as normal.
The idea is that as you use your system a lot, it figures out what you load, and the things you load the most are on the SSD. If you can only fit 5% of your data on the SSD, but that accounts for 50% of hard drive accesses, then half of the time you get the speed benefit of an SSD. And the other half, you have to put up with a slow hard drive. Until you defragment your hard drive, at which point, it completely loses track of what needs to be cached and has to start over, so you get hard drive only speeds for a while. And if you never defragment the hard drive, then the hard drive accesses end up getting really, really slow.
A cache SSD can also be used as a write cache. The idea here is that when you save something to the hard drive, it writes it to the SSD immediately, and then tells the OS that the write is done and it can move on. And then later, it gets around to writing it to the hard drive when it has time.
The main reason to get an SSD as a cache drive is that people who aren't sophisticated enough to understand installing programs on the SSD and bulk data on the hard drive can have software automatically kind of do that for them. Some will argue that if all you can afford is a small SSD, using it as a cache drive lets you get more benefit than if all you could install is the OS. I disagree, but can understand it if all that you can afford is a ~30 GB SSD. But if you can afford 60 GB (which can be had for as little as $50), you're better off using it as a system drive and prioritizing what you want on the SSD.
And if you've already got a ~240 GB SSD? There's no plausible reason for you to consider using an SSD as a cache drive in a desktop or laptop. A cache drive in a system with a large budget could make sense for certain server workloads (and you can actually get them built into a RAID controller), but that's not consumer use.
Thank you! My ssd is 256, and the only thing on it is windows7. Everything else goes on the hard drive, and that only holds current games (gw2, bf3,LotRO), so it's got plenty of space on it. I've never defragged the ssd of course, and I haven't been able to find the command that works with Crucial to do the ssd-type defrag. Not sure if it's done automatically on my ssd.
Thanks for explaining the ssd cache system! Sounds like I don't need to worry about doing that.
edit- oops forgot i did install gw2 to the ssd, but i still have 159g free.
No i will not spend 80$ on an SSD with very limiting MTBF and lower perfomance, to me that would be 80$ down the drain TLC still will not bring you to real high-cap drives for cheap.
For my main drives i have 15K and 10K RPM SAS drives in RAID 0(i use 0 since i do a full stripped image backup every week any how), 2x160GB for OS and critical applications, and 2x320GB for games and such, they still out-perform just about any SSD there with much higher reliablity. For my long term data storage i use regular high cap SATA drives(3x2GB) atm.
I also have a small mini PCIE 16GB SLC HDD which i scavanged from an "old"(4months old) work laptop which i use for SWAP.
SAS drives had, and somewhat still are faster on sequential reads which for gaming is pretty much the most important thing, random reads, and combined IO is only important for general OS operations or specific high perf applications(e.g. databses).
I did bought an Intel X-25 at the time, 64GB for 1K US, did my benchmarks with 85-90% loaded drive at the time and the performance was just bad, even with the price drops on SSD which also brought lower performance especially in the mid-range SSD's it's again not worth it, it would still cost me 2-3K US to switch from regular magnetic SAS drives to SSD's.
FLASH memory in general is comming to it's end, inductive circuts have much higher minimization barriers than conducted ones, for FLASH to evolve you need to start fabricating complex compositve nano materials with high induction index and high tollerance, and it's just not plausible not today, and not in the forisable future, it's physics not engineering.
On the otherhand memristor's are evolving rather well, and unlike NAND flash which is based around a single technology(floating gate transistors), there are currnelty 3(or well 5 since spin based MR's have are split into 3 sub-groups) commercially viable technologies to make them, which means more competition and more possibilities which result in better commercial viability.
So if all goes according to plan and the first commercial MR based chips will be out in mid 2013, there is a good possibility that i will never have to spend 80$ on 1980's tech..
Don't be afraid to install other stuff on the SSD. You probably have a number of small programs that take very little space, so that their space requirements basically amount to a rounding error. You might as well put those on the SSD to benefit from the speed. Web browsers also see a lot of benefit from being on an SSD, as they involve a lot of small file reads and writes.
There's no need for defragmenting an SSD. The reason defragmentation was invented in the first place is that it takes a long time for a hard drive to stop reading, physically move everything to a different spot, and start reading again. The idea is that if you move the pieces of a file around so that they're all together, then the hard drive can at least read a complete file without having to stop to reposition. But even in the extreme corner case where all files are broken into 4 KB clusters that are scattered randomly about the drive, an SSD will still be fast (tens of MB/s, as compared to under 1 MB/s in some tests for a hard drive).
Nope, you can keep on spending $300+ on those 15k rpm drives based on 1950's tech.
And you need to define "outperform" - there are a lot of different metrics.
But for most people, the most important metric is Random I/O. Not even you 15k drives in RAID 0 are going to beat a basic consumer level SSD in Random I/O.
Here's a benchmark from a Seagate Cheetah 15k 300G SAS (actual data from a newegg user):
4KB Random Read: 34,747 IOPS
4KB Random Write: 33,896 IOPS
Max Read: 163 MB/s
Max Write: 166 MB/s
Not bad at all. Here's the Samsung 840 Pro (from Anandtech), mentioned in this article:
Random Read: 100K IOPS
Random Write: 78K IOPS
Sequential Read: 540MB/s
Sequential Write: 450MB/s
Even the least expensive TLC SSD's have an MTBF that will generally give them about 3-5 years of useful life. The MTBF may be different, but guess what, all drives regardless have a rough lifespan of about 3-5 years.
Just to put some numbers to it:
Samsung 840 Pro: 1.5M hours MTBF, 5 year warranty
Seagate Cheetah SAS: 1.4M hours MTBF, 5 year warranty
Also nice to factor in, Energy Use (which translates to heat)
Samsung 840 Pro: 3W peak, 0.3W idle.
Seagate Cheetah 15k SAS: 10W nominal, 13-16W average
Also, sound:
All SSDs: 0dB
Seagate Cheetah 15k SAS: 36dB
Shock (not so important in a desktop, but still pertinent)
Samsung 840 Pro: 1,500G (no moving parts, basically what would crack the PCB)
Seagate Cheetah 15k SAS: 25G operating, 300G parked
Pricetag, with the caveat that the 840 is a brand new drive, and prices have yet had a chance to stablize (this is it's release retail price). Other SSDs based on other technology can be found much cheaper (although the same can be said for hard drives, the Cheetah is probably the cheapest 15k RPM SAS drive you can get, they escalate very steeply from there)
Seagate Cheetah SAS 300G: $209 (with $20 off) - x2 if you want to RAID0
Samsung 840 Pro 256G: $279 (estimated release price, not widely available yet)
You're doing it wrong. Unless you have a very unusual usage model (e.g., mostly spend your time recording videos), a single good SSD would dramatically outperform your RAID arrays. 15K RPM hard drives cost about as much per GB as SSDs, so you're paying SSD prices without getting SSD performance.
Sequential read and write speeds simply don't matter very much for most purposes. Or rather, they matter, but basically everything, including lowly 5400 RPM laptop hard drives, is fast enough that increasing sequential speeds even more barely matters.
The bottleneck that hard drives have is random reads and writes. Rather than getting 100 MB/s in sequential reads, a hard drive may be unable to do even 1 MB/s in random reads. Using 15K SAS hard drives reduces this a little bit--but only a little. RAID 0 doesn't help much with random read times, either. But a good SSD beats a hard drive here by about two orders of magnitude.
And yes, small file random reads are a very common usage model. In fact, that's what you're doing right now: web browsers have to read and write a ton of small files. Most games have a ton of small files that they need to load, too.
-----
Maybe NAND flash will stop scaling with something analogous to Moore's Law sooner than various other things. But it's commercially available today, and works today. It's not some pie in the sky, maybe it will work eventually technology.
If SSDs are 1980s technology (I'm guessing that you're referring to when NAND flash was invented), then hard drives are 1950s technology.
The random Writes are handled by the RAID controller, which has 1024MB DDR2 as main memory for cache. The SAS drives are fast enough for random reads on their own. And again with RAID0 i get about the same sequential speeds as SSD's at a fraction of the cost, the SAS drives were bought when the then brand new Intel X-25 cost over 900US... they come with a 7 year warranty, and onsite next buisness day service from the local segate rep.
The SAS Drives mainly serv for my VMM's and for table swap files.
Also the MBTF for Mechanical Hard Drives and SSD's is really not comparable directly, mechnical drives usually have a MBTF which is based on spin time and the major life expectancy killer for them are spin ups and down i.e restarts, with SSD's every IOP takes a toll on the drive, the MBTF for SSD's is calculated with unrealsticly low IOP count, about 25-30% of the nominal IOP rate that you're comp has on average. It's the same thing as battery life for mobile devices yeah 30 hours with no WIFI, 3G, and the screen brightness set to 10% @ 10c room temp... Unless you restart your computer about 30-40 times a day there is little you can do to actually shorten the MBTF of a mechanical drive these days, with an HDD the more you use it the faster it will die.
Fraction of the cost? Really? How much are those 15k drives a piece - and don't forget to factor in the price of that RAID card either.. because most SSDs are just using onboard SATA3.
And let's look at today's SSD prices, the X-25 hasn't a main stream SSD for a good while now, is listed as Discontinued by Intel, and I can't even buy them new any longer... I quoted the retail price of the Samsung 840 above, and it's considered a pricey SSD by today's standards.
I'm going to leave the rest of your post alone for now, except to say that SSDs are limited in write cycles, you can read from them all day long (so just saying IOP is false - only write ops affect the load count), and that write cycle limit is extremely high when you take into account things like load averaging and excess capacity.
SSD are really nice piece of technology., but currently they are dead-end technlogy. Future SSD's replacement will have to be done using diffrent technology are there are quite a few getting ironed out already for some years.
Of coourse I am talking about ~8-10 years on consumer market, so I am not disencouragining anyone to buy SSD.
I myself I am planning to get one.
PS. Quizz - DRAM and some of other technologies don't suffer from what NAND is suffering. In example they don't have finite number of writes. Of course that does not mean that they are unlimited, but technology is not pre-defining it.
There are two things - first many of other technlogies are way more expensive. Secondly Fabs to make low processes are finite and there is usually more demand for low mn fabs than there are avabile, so prototype / new / niche technologies are usually made & tested using like 120 mn + technologies from really long ago that increase their price even more for those few things that are actually avabile to buy.
So SSD's are great piece of equipment and there is still room for improvement and seriosuly I think many people should consider to get one - in medium-long run - it is dead end and we need other technology.
I'm strongly skeptical of the IOPS numbers there. That's 4 KB "random" read/write speeds not far shy of sequential speeds, which shouldn't be physically possible for a hard drive. There are some benchmarks out there that claim to be "random" reads and writes, but don't often manage to make the hard drive stop and wait for the drive heads to move elsewhere, the way that real programs commonly would.
Here's a review of the same drive:
http://benchmarkreviews.com/index.php?option=com_content&task=view&id=375&Itemid=60&limit=1&limitstart=5
They found 4 KB random read IOPS of 181 and random write of 391 using HD Tune Pro, and 119 and 118 using IOmeter. You can get different numbers depending on the queue depths, and in the harshest tests, even the best SSDs top out around 10k.
DRAM has the big drawback that it requires power to maintain data. That means if your laptop battery were to go dead (and a DRAM drive would be a constant load, even while shut down), that you would lose the contents of your drive. That is why SSDs use NAND, so they can retain data even while powered off.
Here's a review of the same drive:
http://benchmarkreviews.com/index.php?option=com_content&task=view&id=375&Itemid=60&limit=1&limitstart=5
They found 4 KB random read IOPS of 181 and random write of 391 using HD Tune Pro, and 119 and 118 using IOmeter. You can get different numbers depending on the queue depths, and in the harshest tests, even the best SSDs top out around 10k.
Yeah I am fairly skeptical of it as well, but figured I'd use the best numbers I could find out there. They are not from any source of repute:
http://www.newegg.com/Product/Product.aspx?Item=N82E16822148538
However, even using these extremely rosey numbers, they don't compare with an SSD, which was my point.
Yeah I know. It is volatile memory.
Check out things like MRAM or PRAM. They don't need power and are like flash in this matter. There are many others as well. Even wiki will give you basic idea. Most of them have some drawbacks though either technological, patent ones or diffrent.
Maybe so. But every new process node shrink for, say, x86 or ARM processors or DRAM or whatever brings a bunch of new stuff and isn't just exactly the same as the previous except smaller. We don't really know how far they'll be able to extend NAND flash before it becomes impractical to shrink it any further. Nor do we know if or when phase change memory or memristors or whatever will become a commercially viable replacement. OLED monitors have been just around the corner for quite a while. Voice recognition was almost good enough 15 years ago, has come a long way since then, and is still... almost good enough. Stereoscopic 3D has been the wave of the future since the 1950s.
I know that's why I said I am not bad-mouthing SSD's and that I am not against SSD and that I am going to buy one myself soon.
Still it is known already that it is not only SLC -> MLC and not TLC that affect reliabliity and propably 'life lenght' of NAND.
Smaller process nodes ALSO decrease it dramatically.
We also know that in DRAM and some of other technologies smaller process nodes don't affect it that dramatically. Of course that does not mean it might not start to affect it in future - but afaik there are no technological or physic reasons that it should. At least not as dramatically how it affect NAND cells.
I worked on implementing a voice recognition system from an "industry leader" about 15 years ago. Maybe a little less than 15 years ago, but whatever. 15 years ago, they were garbage. We were lucky to be able to use the system as a standard call manager. If I understood how the internet worked back then, I would have kept the recordings of all the people who completely lost their s**t trying to get anything useful out of that system. The other day, I used such a system for something...banking maybe? Don't remember. Anyway, it actually worked. The problem isn't the system understanding what I'm saying any longer, the problem is the people who give the system dumb things to say or provide useless menu options. I still don't think the voice systems will catch on with users. They are good enough, but it's just weird talking to a machine. It's even weirder for me, because I know there are probably a few people clustered around a speaker listening to me talking to the machine.
So...that's the long way around to saying I think you're right. Progress for the sake of progress can be made, but that doesn't mean something is going to be a product that people will want to buy.
** edit **
My prediction for near universal SSD adoption will be when you can buy a 256GB drive for $100 retail. I know zip about any technical details that would bring things to that point, but I think that's the point when people can have their OS and most of the apps they'll use on a single drive, and $100 seems to be the price point for wide scale hard drive adoption.
I can not remember winning or losing a single debate on the internet.
@lizardbones
Not voice recognition and a diffrent but a technology in same area.
Well kinda,depends on how you categorize there things.
Text-to-speech. How far that went in last 10 years.
Propably best text-to-speech on market: http://www.ivona.com/en/
I like this Text to Speech better - http://www.xtranormal.com/, but not because it's good. :-)
Here's a sample of what it can do: http://www.youtube.com/watch?v=WQsaUlBQITI&feature=related
For some reason, the medical profession has picked up on this goofy movie maker.
I can not remember winning or losing a single debate on the internet.
The only thing I would add to this thread is this: People who make predictions about storage devices rarely take into account the fact the programs grow in size too. In fact they have a tendancy to grow in size at a rate that is similar to what hardrives/storage devices can hold.
Putting it another way - Just because you only need 200 or 300 gigs of storage today to run the OS and your main programs does not mean that will hold true even 2 or 3 years from now. As an example - MMOS - Age of Conan the download was something like 5 gigs If I remember right - and everyone crapped in their pants that a game could be so big. It was unheard of. Now a typical game is 20 or 30 gigs and growing.
Putting that aside, it seems that SSD's might perhaps be large enough for some mainstream usage as a performance booster. At the current rate of growth though they will not replace actual Hardrives for a long time to come as people still have movies, half a dozen games, pictures, sound files etc to store and they need a lot of storage capacity to do this. Soon movies and shows are going to transition from HD quality to 3D quality - yet another huge jump in file size.
The last point being - most people dont need the performance that an SSD drive can give and therefore will stick with platter technology which is cheaper, more reliable(long term) and holds significantly more. I do see that SSD drives will start to make their way into the market as an addition to the platter drive in much the same way RAM was added as a cache etc. Not so much as a cache per se but as a place to run the OS and such things as quizzical described earlier, however, the average user doesnt have a clue how to set something like that up - so even those uses will be only a small segment of the population I think.