The OCZ Vertex 3 Review (120GB)
by Anand Lal Shimpi on April 6, 2011 6:32 PM ESTThe Real Issue
While I was covering MWC a real issue with OCZ's SSDs erupted back home: OCZ aggressively moved to high density 25nm IMFT NAND and as a result was shipping product under the Vertex 2 name that was significantly slower than it used to be. Storage Review did a great job jumping on the issue right away.
Let's look at what caused the issue first.
When IMFT announced the move to 25nm it mentioned a doubling in NAND capacity per die. At 25nm you could now fit 64Gbit of MLC NAND (8GB) on a single die, twice what you could get at 34nm. With twice the density in the same die area, costs could come down considerably.
An IMFT 25nm 64Gbit (8GB) MLC NAND die
Remember NAND manufacturing is no different than microprocessor manufacturing. Cost savings aren't realized on day one because yields are usually higher on the older process. Newer wafers are usually more expensive as well. So although you get ~2x density improvement going to 25nm, your yields are lower and wafers are more expensive than they were at 34nm. Even Intel was only able to get a maximum of $110 decrease in price when going from the X25-M G2 to the SSD 320.
OCZ was eager to shift to 25nm. Last year SandForce was the first company to demonstrate 25nm Intel NAND on an SSD at IDF, clearly the controller support was there. As soon as it had the opportunity to, OCZ began migrating the Vertex 2 to 25nm NAND.
SSDs are a lot like GPUs, they are very wide, parallel beasts. While a GPU has a huge array of parallel cores, SSDs are made up of arrays of NAND die working in parallel. Most controllers have 8 channels they can use to talk to NAND devices in parallel, but each channel can often have multiple NAND die active at once.
A Corsair Force F120 using 34nm IMFT NAND
Double the NAND density per die and you can guess what happened next - performance went down considerably at certain capacity points. The most impacted were the smaller capacity drives, e.g. the 60GB Vertex 2. Remember the SF-1200 is only an 8-channel controller so it only needs eight devices to technically be fully populated. However within a single NAND device, multiple die can be active concurrently and in the first 25nm 60GB Vertex 2s there was only one die per NAND package. The end result was significantly reduced performance in some cases, however OCZ failed to change the speed ratings on the drives themselves.
The matter is complicated by the way SandForce's NAND redundancy works. The SF-1000 series controllers have a feature called RAISE that allows your drive to keep working even if a single NAND die fails. The controller accomplishes this redundancy by writing parity data across all NAND devices in the SSD. Should one die fail, the lost data is reconstructed from the remaining data + parity and mapped to a new location in NAND. As a result, total drive capacity is reduced by the size of a single NAND die. With twice the density per NAND die in these early 25nm drives, usable capacity was also reduced when OCZ made the switch with Vertex 2.
The end result was that you could buy a 60GB Vertex 2 with lower performance and less available space without even knowing it.
A 120GB Vertex 2 using 25nm Micron NAND
After a dose of public retribution OCZ agreed to allow end users to swap 25nm Vertex 2s for 34nm drives, they would simply have to pay the difference in cost. OCZ realized that was yet another mistake and eventually allowed the swap for free (thankfully no one was ever charged), which is what should have been done from the start. OCZ went one step further and stopped using 64Gbit NAND in the 60GB Vertex 2, although drives still exist in the channel since no recall was issued.
OCZ ultimately took care of those users who were left with a drive that was slower (and had less capacity) than they thought they were getting. But the problem was far from over.
Facebook
Twitter
RSS
153 Comments
View All Comments
kmmatney - Thursday, April 7, 2011 - link
The answer is pretty easy, I think. Anand's own storage bench is a great test of real world performance, especially the "typical workload"http://www.anandtech.com/show/4256/the-ocz-vertex-...
The bottom line: Version 3 is better than Verion 2, although not by an amazing amount
sunbear - Thursday, April 7, 2011 - link
"3) Finally, are you willing to commit, publicly and within a reasonable period of time, to exchanging any already purchased product for a different configuration should our readers be unhappy with what they've got?"The problem is that it is not straight forward for a customer to know "what they've got" without opening up the SSD and voiding their warranty. OCZ provides the "OCZ Toolbox" that tells you whether your SSD contains 32Gb or 64Gb NAND chips but they don't currently provide any too; to determine whether you have the dreaded Hynix flash or the superior IMFT flash.
I asked in the OCZ forum and their response was to do a secure erase and run the AS SSD benchmark. I have no idea what numbers from the AS SSD benchmark would indicate Hynix versus IMFT.
cptcolo - Friday, April 8, 2011 - link
Hats off to both Anand and OCZ for fixing the Vertex 2 issue. I am really impressed by both Anand and Alex Mei. Anand thanks fo rbeing proactive and presenting OCZ with the problem, and thanks to Alex Mei and Ryan for taking care of the problem 100% (via the change in name and SKUs). You are both true alturists.B0GiE - Friday, April 8, 2011 - link
I just cancelled my order of the 120Gb OCZ Vertex. It says on Scan webpage that it is 550mbs Read and 500mbs Write.Due to this review i'm not sure i believe it. I will wait for further reviews before I purchase a new SSD.
I am interested in game load times for the Vertex 3 such as Black Ops but Anandtech does not show any???
gietrzy - Friday, April 8, 2011 - link
I've just cancelled 120GB Vertex 3 drive. I have no time to investigate whether or not my drive performs as promised.I also have a Vertex 2 60 GB I think "E" version - how do I check if it's faulty.
My scenario is #2 at this page http://www.anandtech.com/show/4256/the-ocz-vertex-...
I also have lots of 1080p avchd videos and even more raw files from my camera so I think I will wait for Intel 510 120 GB review and buy Intel.
One thing's for sure: I will never buy OCZ again.
Thanks Anand, thanks guys!
mattcpa - Friday, April 8, 2011 - link
I ordered the 120GB Vertex 3 from Computers4Sure on the morning before you published this review... :(I also picked up an HDD Optical Bay to put my MBPro 750GB HDD there and plan to put the 120GB in the 6gbps SATA.
I use the Macbook Pro 15" 2.2 SBP for laptop DJ work along with handbraking movies and such; sprinkle in some random gaming.
Hopefully for these processes, it appears this drive will still be near the top of the pack in terms of performance, as I feel I perform many read functions daily rather than performing constant writes. If someone has an opinion, let me know if I am wrong...
Affectionate-Bed-980 - Friday, April 8, 2011 - link
Come on. You HAVE to compare against last generation's Vertex 2. It's selling for $169 at Newegg, and you don't even bench against that. Sigh. Like it's fine if you miss out on some of the other ways say the Kingston, but to skip on the Vertex 2 is a major /facepalm.Shark321 - Friday, April 8, 2011 - link
Yes, Vertex 2 and Agility 2 benchmarks compared to Vertex 3 would be really helpful here.db808 - Friday, April 8, 2011 - link
Hi Anand,First, let me join in with the others in complementing you on your excellent article.
I saw some interesting data hidden in the information describing the IO access patterns of your new IO benchmarks. I was very surprised that the IO size was so small, and that you mentioned that a majority of the IO was sequential.
Some of this can be explained by the multi-threaded nature of the tests. Two applications, each doing sequential IO, running against each other, result in interleaved IOs going to the disk, with a result that is very non-sequential. Some of this may be explained by the application runtime actually requesting 4kb IO, and Windows not having time to do "read aheads".
Windows does have the capability to do larger-IO than was requested by the application (opportunistic IOs), as well as read-ahead and write behinds(that are often coalesced into larger IOs) ... but SSDs may actually be so fast, that the Windows IO optimization algorithms don't have enough time to "think".
You also pointed out that SSD IO performance increases very quickly has the IO size increases above 4kb. It appears that most of the modern controllers parallel stripe the IO across multiple channels, wear-leveling notwithstanding. So an 8kb IO is 2 parallel 4kb IO, for example (ignoring SandForce compression behavior).
The simplest way to cajole the large share of 4kb IO's to 8kb or larger sizes is to simply increase the NTFS cluster size. This has been a performance optimization techniques used with high performance storage arrays for many years. Many Unix systems actually default to 8kb or larger block sizes, and EMC internally uses a 32kb block size as examples.
There is a small negative tradeoff ... some additional slack space at the end of every file. The average slack space per file is 1/2 the cluster size, or 2kb for the default 4kb cluster. Increasing the cluster to 8kb, increases the slack space to 4kb per file ... for a 64kb cluster, it would be 32kb slack per file. The JAM Software "Treesize" utility will actually compute the total slack space for you. With TreeSize Pro, you can even do "what if" analysis and see the impact of changing the cluster size on total slack space.
In summary, slack space overhead only represent a few percentage points of the disk capacity. For example, on my business laptop, by C: drive has about 262K files, and my total wasted space is ~ 644 MB. Increasing the cluster size to 8kb would roughly double my wasted space ... an additional 644MB. Not much.
On my hard-disk based systems that are also memory rich, I regularly run NTFS cluster sizes of 8kb and 16kb ... 64kb for temp file systems. I am pro-actively trading a few percentage points of disk space for higher performance levels. The cost of a few GB of extra overhead on a 1TB disk is a no brainer.
But SSDs are a lot more expensive, and space is a lot tighter. I use a SSD as a boot disk on one PC, and I've filled it about 1/2 full, with the OS, applications, page, hibernate, and temps. Performance is great, and the 40%-ish free space is a form of over-provisioning.
My performance was so good, I had not yet experimented with increasing my cluster size, because I was not able to quantify what the IO size profile looked like. Your IO size statistics from your IO storage benchmark was very enlightening as it shows the (unexpected) large amount of small IO.
On Sandforce-based SSDs, the controller would compress away all the slack space at the end a file, since Windows pads the last cluster in a file with zeros. So with a larger cluster size, your file system would look fuller under Windows, but all the extra slack space would be compressed on the SSD ... with little detriment to the over-provisioning headroom.
I know you are exceedingly busy, but it would be extremely interesting to be able to re-run your controlled test environment with the Anand IO Storage 2011 tests on systems that were built with different cluster sizes. I suspect that using a larger cluster size would improve performance on all SSDs, with SSDs with weaker performance showing the most relative gain. From what I have read, increasing the cluster size beyond 16kb (for Sandforce controllers) will have diminishing (but still positive) returns.
Increasing a Windows 7 boot disk's cluster size from 4kb to 16 kb would increase the wasted space about 4-fold. On my system that would be less than 3GB. It could be a worthwhile trade for performance.
Another reason to explore larger cluster sizes is the fact that the new 28nm Flash chips typically have page sizes of 8kb, not the smaller 4kb used in the 32/34 nm Flash chips. When Windows does 4kb IO on these new 28nm Flash SSDs, it is actually doing sub-page IO, causing the controller to perform a read/modify/write function, and increasing the write amplification effect. The impact would be similar to doing 2kb IO on the SSDs with 4kb page sizes.
If you assume that the typical compression factor is 2:1 for Sandforce controllers, a 16kb NTFS cluster would often be compressed to fit in a single 8kb page ... sounds like a sweet spot.
Using a larger cluster size, also decrease the amount of work needed to append to a file, as fewer clusters need to be allocated. The cluster size also defines the lower limit of contiguousness. This could be important on SSDs, since we normally don't run defrag utilities on SSDs, so we know that fragmentation will only get worse over time.
I will point out that using larger cluster sizes may increase memory usage for the kernel buffer pool, and/or reduce the effective number of buffers for a buffer pool of a given size. I only recommend increasing cluster sizes on systems in a "memory rich" environment.
Again, thank you for your excellent report. Exploring the impact of larger cluster sizes, especially on 28nm based SSDs could add an additional dimension to your analysis. 8kb and larger cluster sizes could further improve real-world SSD performance, and mask some of the performance drop from using the 28nm chips.
db
mpx999 - Sunday, April 10, 2011 - link
That's a big limitation for number of total I/Os. Eg. in 300MB/s SATA-II you'd be limited to ~37.5k IOPS with 8kB transfers, less than some SSDs are capable of, while the limit with 4kB clusters is 2 times higher, which is still beyond current SSDs for random transfers.4kB clusters are a perfect match for x86 processors that use hardware 4kB page size, as each page size is one block on disk. This is especially important for pagefile reads, which tend to be random by nature, rather than pagefile writes that are mostly sequential dumps of memory content. Some Unix systems may use 8kB disc block sizes because it's a default page size for SPARC and Itanium processors. For Power and ARM 4kB is the default but also 64kB can be used. So I'd advice against using large (larger than hw. page size) block sizes on a system/boot partition.
8kB disk block sizes can be useful on partitions dedicated for SQL Server as default database blocks are 8kB for SQL Server, so it's doing 8kB transfers anyway. Oracle supports multiple page sizes and their advice is:
http://www.dba-oracle.com/t_multiple_blocksizes_su...
"Oracle recommends smaller Oracle Database block sizes (2 KB or 4 KB) for online transaction processing (OLTP) or mixed workload environments and larger block sizes (8 KB, 16 KB, or 32 KB) for decision support system (DSS) workload environments."
32kB cluster sizes are a default value on flash cards for digital cameras, as sequential writes of large prictures are done on them.
BTW. The slow speed of both Hynix and Intel 25nm versions of Vertex 2 may be because it's aging controller cannot deal with 8kB flash pages.