Slashdot Mirror
Intel 34nm SSDs Lower Prices, Raise Performance
Vigile writes "When Intel's consumer line of solid state drives were first introduced late in 2008, they impressed reviewers with their performance and reliability. Intel gained a lot of community respect by addressing some performance degradation issues found at PC Perspective by quickly releasing an updated firmware that solved those problems and then some. Now Intel has its second generation of X25-M drives available, designated by a "G2" in the model name. The SSDs are technically very similar though they use 34nm flash rather than the 50nm flash used in the originals and reduced latency times. What is really going to set these new drives apart though, both from the previous Intel offerings and their competition, are the much lower prices allowed by the increased memory density. PC Perspective has posted a full review and breakdown of the new product line that should be available next week."
Fortunately I got it for only about ~$300 so I only "lost" $100 with the new ones coming out. That having been said, I don't regret the purchase at all, it is insanely faster than any other laptop drive out there, while being completely silent and power-friendly. As for TRIM support, I've heard that Intel is not going to add it for the older drives, but I'm not sure if that is just speculation or if it's been officially confirmed by Intel (Intel not expressly say the old drives are getting TRIM support is not the same as expressly denying the support). Fortunately, the drives with the newer firmware don't seem to suffer from much performance degradation, so I'm not really obsessed with TRIM anyway.
Oh and yes, it does run Linux (Arch 64-bit to be precise) just fine.
I can't wait for next year with the ONFI 2.1 FLASH chips (the new drives are not using the new ONFI standard yet) as well as 6Gbit SATA support. At that point I'll put together a new desktop that only uses SSDs, and turn my existing desktop into a 4TB RAID 1+0 file server to handle all the big files... the perfect balance of SATA & spinning media.
AntiFA: An abbreviation for Anti First Amendment.
Getting the prices lower is definitely a move in the right direction. I'm looking forward to moving to SSD in the near future, and not having to worry about hard drive crashes anymore.
You see? You see? Your stupid minds! Stupid! Stupid!
While hard drives will continue to live on for a good while yet where $/GB considerations are paramount (especially archival type applications), the performance advantages of flash drives will soon trump the decreasing cost advantage both for workstation (x25-m) and server (x25-e) environments. The case for flash in servers is even more compelling, where we measure drives in terms of IOPS and a single Intel flash drive performs 10 or 20 times better than the best hard drives on the market for a fraction of the power consumption. Understandably, many IT managers are cautious about adopting new technologies, especially when the failure characteristics are not completely known, but I suspect the advantages are so great that minds are going to start changing, quickly.
The last page of the review states that these should cost you roughly $3 per GB. Whether that's "laughably expensive" depends on what you want to do with the drive.
AnandTech has a nice writeup too. If the price curve drops like the first-gen X-25M we should all be happy pretty soon.
While SSD may be the new kid on the block and show signs of superiority. Hard drives retain a bit of advantage over their non-moving, solid state counter parts. Hard drives can take more write overs than SSD. Flushing the cache to the actual media is still faster on HDD than SSD. SSDs are still very susceptible to static discharge versus HDD due to more surface area having sensitive parts.
I do agree with the parent. SSD are a big thing and they have some important advantages. However, let's not go putting the cart in front of the horse and say that the era of SSD is here upon us. Cost, durability, performance, and longevity are some important areas where SSD needs improvements. In some departments of each of those categories SSD wins hands down. But SSD doesn't win enough in those areas to justify the incredibly high price of the drive. So it is a bit premature to start waving the banners right now.
..and it is fantastic. This was the largest performance increase i've seen on computers in over a decade. I was going to go with a Velociraptor because I knew how important drive access latency was but then Intel patched the fragmentation issue that was worrying me.
I got mounting rails to fit the drive into my desktop case so i'm using it as my primary desktop drive for OS, some applications (Adobe Design Premium Suite runs great on it! Photoshop CS4 loads in 3-4 seconds!), and my main games. I then have a 1.5 TB secondary drive to store my data and music collection etc. I paid around $430 for my 80GB Intel X25-M so being able to get the 160GB for that same price is a fantastic improvement. I will definitely only be going SSD in my machines from now on. Everything loads faster, I get consistently fast boot times even after months of usage.
It is amazing to see Windows XP load up and then all of the system tray apps pop up in a few seconds. You can immediately start loading things like e-mail and Firefox as soon as the desktop appears and there is no discernible lag on first load like you will get with SATA drives since they are still trying to load system tray applications.
"To strive, to seek, to find, and not to yield." - Tennyson
Last year when the x25-m first came out the 80 gig version cost $595, or just a little less than $7.50/gig. Now the same 1st gen drive costs $314 with a -10 dollar discount and free shipping on newegg, or about $3.92/gig.
The new 2nd generation drive 80 gig drive sells for $225, or $2.81/gig. If it follows the same price trend as the 1st gen model around this time next year it should cost ~125 dollars, or about $1.53/gig.
Here are the quick results of the xbench of my 5400rpm 160gig drive in my two year old macbook pro:
Compare those to the results of the new drive here: http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=3607&p=4
Sequential read on the SSD is over 6x faster, and sequential write is 2x faster, but for the performance where it matters the difference is much more noticeable. Random read on the SSD is nearly 140x faster, and random write is over 40x faster.
Couple that performance difference with the lower power consumption, lower noise, and higher threshold for damage, and its a no brainer as to what is the single most price-efficient possible upgrade you can make to a laptop to boost overall performance, responsiveness, and battery life.
I wish I could justify buying one now, but I can't. However, 12 to 18 months from now I will probably be shopping around for a new laptop, and when I do I won't be settling for anything but a SSD. The benefits are just to great to ignore.
How can reviewers be impressed by reliability when they've only had the units for, at most, a year? When these things hit the five-year mark running perfectly well with no data loss in the home/work environment, then I'll be interested.
Ok, they may have been stress tested in factories by the manufacturers, but reviewers don't do that sort of work.
If he's the Walrus then can I be a penguin please?
All you'd need to do to demonstrate to me the greater reliability of an SSD is drop it and a regular hard drive onto the table a couple of times while they're running and see which one keeps running. That would be enough to get me impressed by increased reliability. Regular hard drives are delicate beasts.
This has been covered many times. It's a good number. I can't recall the article, but basically if you write 20GB per day, you'll get more than 5 years out of it thanks to wear leveling and extra space (SSDs actually have more capacity than they make available to you). Now, you might scoff at that but:
1) 20GB/day is a lot for the typical user.
2) People who routinely do more than 20GB/day probably need a lot more storage than SSDs currently provide (you are talking about filling the drive in 4 days) so you probably won't be using an SSD for those purposes anyway
3) People who buy into SSDs at this point in time are typically more on the cutting edge, and are likely no have moved on before the drive wears out.
4) When the drive finally does start having problems, my understanding is that it won't just fail and you'll have lost data. The failure should happen on write, and if it fails to write that will be detectable. If it writes successfully, then it should be readable. If it does fail, I believe that part will just be marked inaccessible and the data will be written somewhere else. The drive should (again, as far as I know) provide details of the failure to SMART and other disk utilities, so the problem can be detected before it progresses to a critical stage. This is much better than magnetic media, where the typical failure is that you go to read data and it is suddenly inaccessible.
Of course, this is all just what I've read about previous generations. I have no data about the 34nm, but I have no reason to suspect it's any worse.
PS. If you want to know how much you currently write to disk and you run a linux system, check out /proc/diskstats. The 10th column should be number of sectors written. Each sector is 512 bytes, so take value*512/1024/1024/1024 and you'll get the number of GB each device has written since bootup.
Let's make some wild predictions based on recent price trends. (Trends found here). Over the last few years, flash memory has been increasing in GB/$ at a rate of 185% per year. Meanwhile, hard drives have slowed to only 42% improvement per year.
Based on these trends, here is the estimated cost of 10 TB using either technology:
July 2009: Platter = $750, Flash = $28,125
July 2010: Platter = $528, Flash = $9,868
July 2014: Platter= $130, Flash = $150
July 2019: Platter= $23, Flash = $0.80
July 2024: Platter= $4, Flash = $0.004
In July 2024, a 10 PB flash drive would cost $42! Of course, we can't assume these trends will continue, but it seems a good bet that we won't be worrying about the size of our mp3 collections. The traditional hard drive may only have five years of competitive life remaining.
Actually, surprisingly, you do need to run a kind of defragmentation.
Just not the usual one.
That's because writing in flash is in pages (4k?) but erase can be done only in blocks of 512k. So what happens is that controller have to do some insane job of joggling your writes and rewrites to spread or combine or whatever... on the fly...
As a result, after intensive use, the address space become fragmented, just like memory heap in regular software after lots of allocations/deletions.
Currently, the only way to restore performance is to issue low level format command - secure erase or some such.
I think AnandTech wrote a big piece on it.
The TRIM thing will help to delay (or even eliminate) the need for such drastic measures.
(Yes, I know the new parts are 34 nm)
I thought the progression of feature size went: 90 nm, 65 nm, 45 nm, 34 nm.
But the graphics processors seem to be using 55, and these SSDs are being reduced from 50.
I thought they had to pour gazillions into standardizing fab construction, steppers, and all the equipment. So is some plant manager stumbling in with a hangover one morning and accidentally setting the big dial for 50 or 55 or something? What's the deal here?
My other car is a 1984 Nark Avenger.
I think the UPS will cover that.
Let's make a few predictions based on recent trends:
July 2007: number of wives = 0
July 2009: number of wives = 1
July 2011: number of wives = 2
July 2013: number of wives = 3
July 2015: number of wives = 4
July 2017: number of wives = 5
July 2019: number of wives = 6
July 2021: number of wives = 7
Gosh, I'll need to implement wear levelling soon, too.
Extrapolation: almost as good as copulation.
The Cloud - because you don't care if your apps and data are up in the air.
>Sequential read on the SSD is over 6x faster, and sequential write is 2x faster,
>but for the performance where it matters the difference is much more noticeable.
>Random read on the SSD is nearly 140x faster, and random write is over 40x faster.
So
>Not random writes, not sequential reads, and not anything not HD-related.
is wrong.
It also seems to me that you don't really need to say
>[no performance increases on] anything not HD-related.
or
>They don't help with anything CPU-, RAM-...-intensive
when you are talking about hard drive upgrades.
And of course it does help with I/O intensive stuff if that I/O is to the HD.
My RAM and CPU speed are fine, but my second upgrade (when I can afford it) will be an SSD (my first upgrade will be a video card - I currently have an Intel x3100, good for bleeding edge Xorg stuff, but low-powered, and Radeon[HD] will be catching up before I can afford it).
P.S. I wish slashdot would quote like a mail client (or a *chan), Also, the preview should not leave out blank lines if they will be present in the final post
Climate Progress - Hell and High Water
Cool, thanks for the tip!
cat /proc/diskstats | grep "[sh]d[a-z] " | awk '{print $10 "*512/1024/1024/1024"}' | bc -l
I suspect they have a capacitor large enough to finish committing their buffers. At least they seem to see little performance degradation with write barriers, and do retain all the files they should when I pull the power while writing. (I didn't do a proper test, but it seems to work correctly, assuming your OS does.)
(And for the record, any OS that still thinks anything the HD acks is written is living in a dream world, it hasn't been true for 15 years on consumer disks.)
Yes, I'm aware of what is in that document (that's how I figured out what the columns were to begin with). That document skips over the first 3 columns of the output for it's numbering (major device number, minor device number, and device name). It considers column 4 to be field 1. Not sure why they wrote the document that way, but PsychiKiller's command above uses awk to print out the 10th column, and that does indeed give you the number of bytes written.
But that IS extending the life. Without wear leveling, if I've got an 80GB drive and I store 50GB of data on it which I frequently modify, then after X years that 50GB will be worn out and I'll be left with 30GB. That isn't enough for me to use, so essentially the drive is dead as far as I'm concerned. Now consider a drive with wear leveling. After X years, I will only have used up 5/8 of the write cycles across the entire drive. I can still use the drive for another 0.6X years. Wear leveling has extended the useful life of the drive by 60%.
But even for more typical usage, it's possible for wear leveling to actually extend the number of writes that can be done if the wear leveling works in certain ways. For this to make sense, you have to understand how SSD storage is organized. Much like a HDD, which is organized into sectors, clusters, platters, etc, we have a similar organization with SSDs. You have bytes grouped into pages, and multiple pages are grouped into blocks (and it goes on from there).
The smallest group of data which you can write on an SSD is a page. However, the smallest group you can erase is a block.
SSDs don't allow you to overwrite a page with your new data. Instead, you must first erase it and then write the new data to it. The problem here is that you have to erase a block at a time, but the rest of the pages in the block could already contain other data. So what happens is that the controller copies all of the pages that you don't want to modify from that block into cache, erases the block, writes back all of the page that are staying the same, and then writes your new block. Now surely you can see the problem here...you've only intended to write to one single page, but you've also used up a write cycle for every single page that you DIDN'T modify.
So how can wear leveling help this? Well, lets say that block consists of 10 pages, and only 9 of those pages are filled. You now want to modify one of those 9 pages. Well, instead of doing an erase, which uses up a write cycle on 9 of the 10 pages, the wear leveling can simply say "OK, I won't erase page 4...instead I'll just remember that I don't care about the data stored there. I'll also write this new data for page 4 into page 10 and remember that the data is now stored there". Thus to make that modification, we only use up a write cycle on a single page instead of 9 of the pages. Now, the next time we go to make a write, we'll have to erase the entire block and write to 9 of those pages. However, we'll once again have an empty page, so on the 3rd write we can do the same thing we did the first time. As a result, instead of a single modification writing to 9 pages each time, it averages 5 page writes each time (alternates between 1 and 9 pages).
Of course the wear leveling can be extended to perform the same type of thing across multiple block. The advantage here would be that, as lots of data gets modified, each page may eventually be migrated out of that block without the block having to be erased. Eventually, we could end up with the block being empty and then we can erase it without rewriting pointlessly to any of the page (or if it's almost empty, we'll only rewrite a few pages).
Other things wear leveling could do is recognize that some blocks never seem to get modified, and then shuffle that data to a different spot on the drive so that you don't end up with certain blocks that suffer almost no write-wear while other blocks are reaching their limit.
I don't know which specific techniques current SSD drives implement, but these are a few possibilities. I'm sure there are others.
The OS thinks that everything ACKed as sent to the storage unit is written,
What does it matter what the OS "thinks"? When power is lost, all of its "thoughts" disappear. When you power it back on it reloads its "thoughts" from the DISK, thus there can be no confusion.