SSD 2010 roundup: 15 SSDs compared - BeHardware
>> SSD

Written by Marc Prieur

Published on September 17, 2010

URL: http://www.behardware.com/art/lire/794/


Page 1

Introduction



At last! After the appearance of our article SSD, TRIM and IOMeter, our SSD product report is ready! On the menu, you will of course find the SSDs that we have been recommending up until now, namely those based on Indilinx Barefoot controllers and the Intel Postvilles. The Intel Postvilles have been tested in 80 GB and 160 GB versions, while the Barefoot is used in no less than 4 of the SSDs tested:

- The Crucial M225 64 GB and 128 GB, which are standard models
- The OCZ Vertex Turbo 120 GB, with an overclocked version of the Barefoot
- The G.Skill Falcon II, which combines a Barefoot with IMFT 34nm flash

Still with a view to give representation to the ďoldĒ SSDs, we have also included the OCZ Summit 60/120 GB based on a Samsung controller and which are clones of the Samsung PB22-Js.

5 new SSDs join these:

- The OCZ Vertex 2 100 GB, based on the already famous SandForce SF-1200 controller
- The Crucial C300 128 GB, using the Marvell SATA 6 Gbits/s controller
- The Sandisk G3 120 GB, long-awaited and finally available
- The Kingston SSDNow V SNV425, built around the JMicron JMF618 controller
- The Kingston SSDNow V+ SNVP325, using a Toshiba controller

What is an SSD?
Before anything else, and for those who arenít necessarily up on the technology, hereís a short recap of what an SSD is. An SSD (or Solid State Drive) is a storage device constituted of flash memory, in opposition to a standard hard drive (HDD) which stocks data on magnetic platters.

SSDs have many advantages over standard hard drives, starting with performance levels and including noise and resistance to knocks. Usage is transparent for the system, they are addressed in the same way as hard drives by the SATA controller.

Their disadvantage comes in the fact that flash memory has a limited lifespan. MLC flash NAND can only be written between 5,000 and 10,000 times depending on its quality but fortunately this is compensated by wear leveling algorithms which share usage across cells and take this drawback completely out of the picture unless you fully rewrite your SSD everyday, which doesnít correspond to standard usage.

The memory chips are also limited in terms of data retention, because while a new cell can stock data for 10 years, at the end of its life this falls to a year. In practice and now the technology has been available for a couple of years, it looks as if the confidence placed in SSD reliability by manufacturers resulting from wear leveling algorithms has been justified. SSDs do not fail any more frequently than hard drives.


Page 2
Indilinx Barefoot

Indilinx Barefoot
First coming onto the market at the beginning of 2009, initially on the OCZ Vertex, the Indilinx Barefoot is based on an ARM7 processor and has 4 flash channels each of which is made up of 4 chips (16 in all). It can work in tandem with SDRAM which is used as the cache, while TRIM support is afforded with firmware dating from October 2009. Along with the Intel SSDs, SSDs based on the Barefoot were up until now our SSD of choice.

In addition to the manufacturers included in this report, many others also have an SSD offer based on the Indilinx: Corsairís Extreme range (with 45nm Samsung flash), Nova (34nm Intel/Micron), and SuperTalent.
Crucial M225 64 GB
Launched in July 2009, the Crucial M225 uses the Indilinx Barefoot which first appeared at the end of 2008 in the OCZ Vertex. A 64 MB RAM chip from Elpida serves as the cache, with the NAND made up of 16 Samsung 4 GB chips engraved at 45nm placed around the SSD.

Crucial M225 128 GB
The 128 GB version of the Crucial M225 uses a different type of flash than the 64 GB one, going for 8 GB rather than 4 GB chips (still 16 in number).

OCZ Vertex Turbo 120 GB
Not as labelled, this SSD offers the same capacity as the M225 128 GB, namely 128 GB and around 119.2 GiB. Launched in July 2009, like the M225 it uses an Indilinx controller, 64 MB of Elpida RAM and 16 x 8 GB 45nm Samsung chips Ė we were able to check this without being able to open the SSD completely, due to a screw that we couldnít remove. In fact, the difference comes in the fact that the controller and memory cache are clocked at 180 MHz rather than 166 MHz by default.

G.Skill Falcon II 128 GB
In November 2009 G.Skill launched the Falcon II, following the Falcon in April. Here the Indilinx Barefoot controller is in whatís known as its ďECOĒ version, which can be used with 34nm Intel/Micron flash. This is the memory used here, in the form of 16 x 8 GB chips, while a 64 MB Elpida chip is again used as the cache.



Page 3
Intel & Samsung

Intel PC29AS21BA0
In July 2009 this new revision of the Intel controller came onto the market Ė the previous one used was the PC29AS21AA0. With 10 flash channels like the previous version, the new one gave improved performance, a 32 MB RAM chip (against 16 MB previously), 34nm flash and TRIM support via firmware in October 2009. Note that these SSDs are also sold by Kingston under its own brand.
Intel X25-M "Postville" 80 GB
In its 80 GB version, the Intel SSDSA2M080G2GC (ouch!), or more simply the X25-M Postville, uses the Intel PC29AS21BA0 controller with 32 MB of Micron RAM and 10 x 8 GB 34nm Intel/Micron flash chips.

Intel X25-M "Postville" 160 GB
A similar design is used as on the SSDSA2M160G2GC, but this time the RAM is from ISSI while the chips are from the same family but in 16 GB versions.

Samsung S3C29RBB01
Announced at the end of 2008 for availability at the beginning of 2009, this new controller, based on an ARM processor like the Barefoot, has 8 flash channels against 4 for the previous model. In tandem with a RAM chip itís used on numerous SSDs such as the Samsung PB22-Js of course, as well as the OCZ Summits, the Corsair Performance Series and the 1st gen Kingston V+s. TRIM support comes via firmware dated end 2009.
OCZ Summit 60 GB
This is a 64 GB SSD based on the Samsung S3C29RBB01 controller, with 128 MB of Samsung RAM and 8 x 8GB Samsung 45nm chips. The PCB is in fact made by Samsung, as you can see from the label on the back, and is equivalent to that on the PB22-J.

OCZ Summit 120 GB
In its 120 GB version, or more exactly 128 GB, the OCZ Summit retains the same design as on the smaller version but with 16 rather than 8 flash chips.



Page 4
SandForce, Marvell & SanDisk

SandForce SF-1200
Unveiled to the public at the end of 2009 on the occasion of a partnership with OCZ (them again!), the SF-1200 is SandForceís general consumer SSD. In contrast to a lot of modern controllers, it doesnít use an external cache but rather a small internal one coupled with sophisticated algorithms grouped under the name DuraWrite. Among the algorithms used is the compression of written data onto the SSD, which has the merit of economising the flash but which naturally wonít work on files that are already compressed, as is the case with archives as well as most multimedia files (mp3s, jpegs, mpegs and so on).

While most SSDs reserve around 6.8% of flash space for wear leveling and internal optimisation, SandForce goes further, reserving 12.6% (60/120/240 GB versions) or 27.2% (100/200 GB versions). Many manufacturers offer SSDs based on the SF-1200, such as OCZ with the Vertex 2s and Agility 2s, Corsair with the Force Series and G.Skill with the Phoenix.
OCZ Vertex 2 100 GB
Launched in April 2010, the OCZ Vertex 2 100 GB uses the SandForce SF-1200 without RAM but with 128 GB of flash memory made up of 16 x 8GB chips manufactured by Intel/Micron at 34nm. This SSD comes with firmware exclusive to OCZ which unblocks the SF-1200 in terms of random writes, taking it from 10,000 IO/s to 30,000 IO/s.
Corsair F120
Announced in May 2010, the Corsair F120 is part of a Corsair range using the SF-1200 and available in 40, 60, 80, 120 and 160 GB versions. It is similar to the Vertex 2 120 but with reserved space limited to 12.6%. The whole Corsair F range goes beyond the 10 000 IOPS limit that is found on some SF-1200 SSDs.

G.Skill Phoenix Pro 120 GB
G.Skill also has two ranges of SandForce SSD, the Phoenix and the Phoenix Pro range. The Phoenix range is limited to 10 000 IOPS in random writes by the firmware but not the Phoenix Pros, as is the case with the OCZ Agility 2 and Vertex 2 ranges. In practice, this sort of limitation has no impact on desktop usage. The SSD uses the now traditional 34nm Intel/Micron memory.

Marvell 88SS9174
Very active when it comes to SATA 6 Gbits, Marvell offers controllers designed for motherboards introducing this technology onto Intel platforms (Intel chipsets are currently limited to SATA 3 Gbits) as well as an SSD controller used for the moment exclusively by Crucial.
Crucial C300 128 GB
First SSD with the Marvell controller, the C300 was announced in December 2009 by Micron and in January 2010 by its subsidiary Crucial. The Marvell 88SS9174 comes with a Micron cache of 128 MB and 16 x 8GB 34nm Intel/Micron chips.

SanDisk SDC4
Long anticipated as it was first presented at the beginning of 2009, namely at the same time as the Indilinx Barefoot, the Sandisk G3 only arrived on the market a year later. Like SandForce, SanDisk reserves more flash than the norm (12.6% against 6.8% for the others).
SanDisk G3 120 GB
In its 120 GB version, the G3 uses the SanDisk SDC4 controller which comes with 64 MB of RAM originating from Hynix and addressing 16 x 8 GB MLC SanDisk chips. Note that SanDisk gives a 10 year guarantee on this SSD, while other manufacturers give 2 or 3 years! At a time when some still doubt SSD reliability, SanDiskís guarantee should be attractive.



Page 5
JMicron & Toshiba

JMicron JMF618
Unfortunately famous for its JMF602 controller, which gave random writes at such low speeds that lags could be felt during usage, JMicron is back with the JM618, a derivative of the JMF612 supporting 34nm Toshiba flash. This time it uses external memory to serve as cache, so as to palliate the lags. This controller is used on many SSDs such as the Kingston SSDNOW V Series, WD SiliconEdge Blues or the Corsair Reactors.
Kingston SSDNOW V SNV425 S2/128 GB
After the JMicron JMF602, the V Series have changed up for the JMicron JMF618, here described as from Toshiba. It comes with 64 MB of Winbond RAM as well as 16 x 8 GB 43nm Toshiba chips. Note that on the SSD casing Kingston has placed a thermal pad designed to cool the JMF618, which may indicate that the JMF618 heats up more than other controllers.

Toshiba T6UG1XBG
Developed by Toshiba, the T6UG1XBG controller is used on their own HG2 SSDs as well as by Kingston, which has a partnership with Toshiba for SSDs. Unfortunately we donít have much information on this controller except that itís engraved at 43nm and supports TRIM natively, like all the latest controllers.
Kingston SSDNOW V+ SNVP325 S2/128 GB
The only manufactuer to go for the T6UG1XBG, apart from Toshiba themselves, is Kingston, who have set it up with 128 MB of JMicron RAM and 8 x 16GB 43nm Toshiba chips.


Note the extensive bundle: a 2.5 to 3.5 inch adaptor, a 2.5 inch external USB casing, an Acronis True Image license, an SATA cable and a Molex to SATA adaptor (the SNV425 has the same but without the external casing)!

In terms of cooling, Kingston goes even further than on the SNV425, with the thermal pad covering the entire SSD casing!



Page 6
Test protocol

The test
As we say in our articleSSD, TRIM and IOMeter, testing SSDs in the right way, moreover in an environment that is compatible with TRIM, demanded a complete reworking of our test protocol. The arrival of controllers that include data compression algotrithms is a new parameter to take into account.

While previously we took readings of SSD access times with the help of h2bench, this reading has been sidelined as it can give false results if the SSD is empty. Certain controllers give similar readings for random accesses as sequential accesses.


Speeds are now only measured using IOMeter, that we have compiled in its latest version so as to be able to play on the type of data read or written on the SSD, so that it may be compressible or not by the SandForce controller. The two types of data are tested on controllers that compress data on the fly so as to give best and worse cases. Speeds are measured in several cases:

- Sequential reads by 2 MB blocks
- Sequential reads by 4 KB blocks
- Radom reads by 4 KB blocks
- Sequential writes by 2 MB blocks
- Sequential writes by 4 KB blocks
- Random writes by 4 KB blocks

This allows us to measure the speed of the storage device in terms of ins / outs (IOs) as well as throughputs. The effectiveness of NCQ, which allows the processing of concurrent commands, is also measured by testing random accesses, read and write, with 1, 2, 4 and 8 simultaneous commands.

Next come the practical tests with writes and reads of various groups of files. These groups are made up of:

- A collection of large files: 6.8 GB on average
- Medium: 796 KB on average
- Small: 44 KB on average


The source or the target when reading or writing to the SSD is a RAID of six 150 GB VelociRaptor drives mounted on an ARECA ARC1280ML PCI-Express x8 card with a strip size of 8 KB.

Lastly there are some purely practical tests, namely various operations timed after installation of Windows 7 64-bit on each of the SSDs:

- Start-up of Windows 7
- Installation of Photoshop CS5
- Start-up of Windows 7 + various applications
- Launch of a Crysis level

As it is halfway between a synthetic test and a practical test, PC Mark Vantage has been abandonned. At first we wanted to retain it but the results on this type of benchmark, like all those based on recording of accesses over the course of one session, are not valid with the SandForce controller. These benchmarks are based on logs which have recorded accesses to be repeated, but not the data contained in these accesses. These means that the data used in the benchmark may well be highly compressible, which isnít necessarily the case in real usage.

The test machine is a Core i7-920 mounted on a Gigabyte GA-X58A-UD3R with 6 GB of DDR3-1333 and a Radeon HD 5870. The synthetic tests are carried out in Windows 7 64-bit with the storage system as a secondary drive, the boot drive being an X25-M 80 GB, all with Intel RST 9.6 drivers in AHCI mode. The practical tests with the test hard drive with the OS are carried out with the Microsoft AHCI drivers from Windows 7. This same driver is used if an SSD is tested at SATA 6 Gbits on the Marvell controller on the motherboard.


Page 7
So then, does it TRIM?

So then, does it TRIM?
As we have already explained on several occasions (including in this article), the TRIM command is important because it enables to avoid the SSD performance deterioration over the course of use.

Of course, the impact on performance varies from one controller to another. By carrying out sequential writes on an SSD on which we have previously carried out random writes, and vice-versa, we reached the following conclusions with respect to the sensitiveness of controllers to fragmentation:

- Indilinx Barefoot :
Random writes may be reduced by a factor of 4.5, sequential writes by a factor of 5
- Intel PC29AS21BA0 :
Random writes may be reduced by a factor of 4, no significant impact on sequential writes
- JMicron JMF618 :
No significant impact on random writes, sequential writes may fall by 10%
- Marvell 88SS9174:
Random writes may be reduced by a factor of 4, sequential writes by 33%
- Samsung S3C29RBB01:
Random writes may be reduced by a factor of 14, sequential writes by 2.5
- SandForce SF-1200:
Random writes may be reduced by a factor of 2, sequential writes by 4
- Sandisk SDC4:
No significant impact on random writes, sequential writes may be down 25%
- Toshiba T6UG1XBG :
No significant impact

The Toshiba and JMicron JMF618 controllers are those that suffer the least fragmentation. This is probably due to the fact that their random write performance is poor which makes us think they employ very little write combining. The Sandisk G3 also does fairly well. Only random write performance on Intel SSDs is affected but this is fairly high in the first place. The Marvell is only slightly affected when it comes to sequential performance, more so for random performance, but again the numbers are fairly high in the first place. Those controllers that are most affected are the Indilinx Barefoot, the SandForce SF-1200, and the Samsung S3C29RBB01, with the last especially affected when it comes to random writes.

Among those affected in terms of sequential writes, some controllers completely recover from fragmentation linked to random accesses after just one sequential write. In particular, this is the case with Marvell, JMicron and Samsung. There are gains with Indilinx and SandForce controllers but they are only partial. Other mechanisms such as Garbage Collector attempt to use times when the SSD is at rest to defragment it, but this type of mechanism brings about random, or even inexistent, gains in spite of whatís announced.

Only TRIM allows you to maintain the initial level of performance, whether this be for sequential or random accesses. To recap, to take advantage of TRIM, youíll need a combination of 3 elements that support the command:

- The operating system (Windows 7, Linux, soon Mac OS X)
- IDE/AHCI controller drivers (in W7, Microsoft or Intel drivers)
- The SSD

We can put the AMD or Marvell drivers to one side because, even in their latest versions they donít support the command: this means you have to ďmake doĒ with the generic Microsoft driver (Update : latests AMD and Marvell drivers are now TRIM compliant !). What about support for the TRIM command by these SSDs? On paper they all support it, but in practice this is far from being the case:


The SSDs based on the Indilinx, Intel and Marvell controllers support TRIM correctly.

This is also the case for the SandForce SF-1200, and therefore the Corsairs and G.Skills not included in the table, even if its implementation is only partial when you use the generic Microsoft driver, obligatory on any platform other than Intel for TRIM support. As Microsoft hasnít followed the latest ATA specifications according to SandForce, support for the TRIM command, which was inexistant before firmware 1.1, is now included but only partially as at best it only acts on 4 GB per sweep, which does however cover the great majority of usages. In time, Microsoft should update its driver which will mean the 4 GB limit will be removed just like when using the SF-1200 with Intel drivers. SandForce has also told us that they are looking into a workaround which would have the same effect.

We give a red card to those SSDs that donít support TRIM in practice when in fact itís announced as being supported! This goes for the Kingston SNV425 and its JMicron JMF618, as well as the OCZ Summit based on a Samsung controller. Nor does TRIM work on the SanDisk G3, though after receiving a first SSD that was defective here, the second seemed to be doing something with the command, given the fact that the quick formatting of one partition was fairly slow (30 seconds to 1 minute compared to 5 seconds without TRIM). For all that, performance levels, though only slightly affected, didnít recover.

Finally, itís impossible for us to determine if TRIM functions well on the Toshiba controller as it suffers from no performance deterioration.


Page 8
Sequential speeds

Sequential speeds
For all the synthetic tests, we used IOMeter, a powerful tool that weíve been using for a long time and that we talked about earlier in this article. We configure the accesses so as to align them to a value of 4 KB and not 512 bytes as is the case by default and which doesnít correspond to the usage patterns of an SSD on a modern OS. The tests were carried out using two methods:

-Without partition for the write tests
-With partition for the read tests

When used on one partition, IOMeter starts by filling this partition with a file written sequentially. This means this isnít a good method for writes because the SSD wonít have enough free blocks to optimise random write performance, which isnít the case in an environment that supports TRIM. Between each write test, we create and carry out a quick formatting on one partition which is then deleted, so as to ďtrimĒ the whole SSD.

For reads, we didnít carry out the tests on an unwritteen SSD because on an unwritten SSD some controllers process random reads sequentially. This is obviously not the case when you have to read real data randomly and we therefore use IOMeter with a partition so as to obtain an SSD filled with data.


Another important point, we use the latest beta version of the software, which we compile ourselves. This allows us to work either with standard data used by most benches and which is easily compressible (series of 0s and 1s), or with random data which canít be compressed. In the first instance, SandForceís DuraWrite can be used, but not in the second where it has to write and read the information bit by bit, or almost. The random data tests for the Vertex 2 are given as ďVertex 2 100 GB/ RĒ in the graphs.


We start with the sequential speeds using 2 MB blocks of data which allows us to reach maximum bandwidth of the SSD. In reads, the Crucial C300 is in the lead with the SATA 3 Gbits and particularly the SATA 6 Gbits, which scores 338 MB/s! The Vertex 2 and other SandForces are hot on the heals of the SATA 3 Gbits card when working with easily compressible data, reaching speeds of 264 MB/s, but this falls to 203 MB/s when the data canít be compressed. Overall, all SSDs offer top drawer sequential read speeds.

The results for writes contrast with these, with the X25-M 80 GB down at just 79 MB/s. Of course, the main usage for an SSD, especially one this size, isnít to house large files, but anywayÖ The Vertex 2 is top of the tree here as long as itís able to compress data; otherwise speeds drop from 253 MB/s to 135 MB/s. The fastest SSD therefore is the Vertex Turbo, followed by the Summit 120 GB and the Crucial M225 128 GB. Note that Samsung flash is faster on sequential writes, with the Falcon II notably slower than the M225 here.


Staying with sequential accesses, we move on to small 4 KB blocks. This will put maximum pressure on the controller in terms of ins / outs. Unsurprisingly speeds are significantly down and itís the Sandisk G3 that suffers most. The Crucial C300 isnít very well placed here when used with the Marvell 6 Gbits controller and the Microsft driver (TRIM wasnít activated with the Marvell driver). It does much better on the Intel 3 Gbits controller and Intel driver. Itís then in joint first place with the Intel X25-M.


Page 9
Random speeds

Random speeds
Itís with random data that SSDs really have the advantage over standard hard drives, as random performance is linked to the ultra fast access times of SSDs.


Once again, note the relatively poor results with the C300 SATA 6 Gbits on the Marvell with Microsoft drivers in comparison to SATA 3 Gbits with the Intel driver. The Vertex Turbo is at the head of the pack for read performance, closely followed by the Crucial C300 and the rest of the Indilinx SSDs. The Vertex 2, based on the SF-1200, is therefore a little disappointing while still excellent compared to magnetic hard drives. At the back of the pack, the Sandisk G3 gives a truly poor performance and the same goes for the SNV425.

In random writes the SandForce SF-1200 is outstanding. Itís closely followed by the Crucial C300 and these two SSDs are far ahead of the competition, the Indilinxs and Intels, whose performance is however nothing to be embarrassed about in the absolute. The SSDs based on Samsung, JMicron and Toshiba controllers donít perform well here and are down on hard drive performance.

Note however that here we give speeds over 5 minutes, after a minute without recording the results so as take momentary and unrepeatable highs in performance out of the picture. Some controllers can give higher momentary maximum speeds, especially those low down in the rankings. On an unused Samsung drive for example, speeds can get up to 3.4 MB/s for random writes, but given that TRIM isnít supported this doesnít continue for long. Over a few seconds the Kingston SNVP325 can get up to 4 MB/s, while the SNV425, based on a JMicron, reaches 19 MB/s for a minute before collapsing.


Page 10
Random speeds & NCQ

Random speeds & NCQ
As on the previous page, here we measure SSD performance during random accesses but at the same time multiplying simultaneous accesses. Thanks to NCQ (native command queuing) and internal SSD optimisations, the performance levels of some SSDs do in fact increase during simultaneous accesses, something that is typical of heavy usage.


In terms of gains first of all, 3 controllers offer very good scaling when it comes to reads: the Intel, the SandForce and the Marvell. The Indilinx sees performance levels climb up to four concurrent comands, against two for the Samsung and SanDisk. Thereís a very slight increase with the JMicron and Toshiba but itís very limited. In terms of absolute performance the Crucial C300 is in the lead, followed by the X25-M and the Vertex 2.


In terms of random writes, gains are harder to produce with NCQ. Only the Marvell and SandForce controllers, and to a lesser extent Intel, show gains. The performance levels you get with the Crucial C300 and the SandForce are quite simply exceptional, with speeds close to those for sequential writes of 2 MB blocks! Note some anomalies in the results with the Corsair F120 and Phoenix Pro, with in particular a fall in performance with 8 commands which you don't get on the Vertex 2. Levels are nevertheless very good.


Page 11
File copying

File copying
This brings us to file copying. We measured read and write speeds obtained when copying various groups of files with Robocopy. These groups are made up of:

- A collection of large files: 6.8 GB on average
- Medium: 796 KB on average
- Small: 44 KB on average

The source or the target when reading or writing to the SSD is a RAID of six 150 GB VelociRaptor drives mounted on an ARECA ARC1280ML PCI-Express x8 card with a strip size of 8 KB. On the VelociRaptor in the graph, the files are read and written on a partition that begins halfway in on the drive.


For reads of medium and large files, the Crucial C300 is again the outstanding performer. With large files the SATA 6 Gbits/s interface even increases its advantage significantly as speeds approach 300 MB/s, a level that can be surpassed with the 256 GB version of the C300. The C300 is ďjustĒ second with small files, the Kingston SNVP325 doing best here. All the SSDs give a good level of overall performance, with the exception of the Sandisk G3 for which performance levels collapse with small and medium sized files.


The Kingston SNVP325 is top of the pile for large and medium sized files but the Vertex 2 is best for small files, which, it should be noted, can be slightly compressed in contrast to large and medium sized files. The X25-M 80 GB is last by some distance for large files, which is logical given the results on the preceeding theoretical tests, but with medium sized and small files the OCZ Summit, Sandisk G3 and Kingston SNV425 bring up the rear by some distance.


Page 12
Practical tests

Practical tests
After raw performance levels and file copying, we now move on to some slightly more meaningful numbers, the practical tests. While for the other tests, the SSD was set up as a secondary drive, here it becomes the primary drive. After installing Windows 7 64-bit we time various tasks:

- Windows 7 start-up
We measure how long it takes to start up Windows 7, from initialising the load to the appearance of the Windows desktop.

- Windows 7 start-up + various applications:
This is the time required to start up Windows 7, Adobe Photoshop CS 5, Excel 2010, Word 2010, PowerPoint 2010 and Outlook 2010.

- Launch of a Crysis level:
We measure how long it takes to launch Crysis and a game level launched straight from a command line.

- Installation of Photoshop CS5:
This is the time required to install Photoshop CS5 from the Adobe site download archive to the SSD. Installation is broken down into two stages, the extraction of files from the archive, then the actual installation.

- Installation of Office 2010:
Here we measure the time required to install Office 2010 from an image on the SSD.

For information, in addition to the VelociRaptor 150 GB weíve been working with since the beginning of the tests, we have also included results on a Samsung F3 1 TB (7200 rpm).


Loading Windows is twice as fast on an SSD as on the VelociRaptor, while loading of Windows + the applications (even heavier) is 3.5 x as fast. The impact on Crysis is more limited. We have to admit that loading a game level is generally more a question of sequential reads and data decompression. There is nevertheless an improvement.

The results of all the SSDs are quite close to each other but the Sandisk G3 120 GB and Kingston SSD Now V 128 GB (SNV425) nevertheless bring up the rear. The other SSDs are very close, with first place going to the Crucial C300. For information, we didnít note any gain in performance when using the SATA 6 Gbits port in these practical tests. The Crucial C300 is closely followed by the X25-M then the Indilinx/SandForce SSDs.


Here the gains are much less significant. The first stage of the installation of Photoshop shows similar results on all the storage devices Ė extraction from an archive is mainly limited by decompression speed, which depends on the CPU and memory. You do however get bigger gains on the second part of the installation process. The installation of Office 2010 is faster on SSDs but the difference with standard drives is only around 10%.

In terms of the SSDs themselves, though the results are relatively close, the SanDisk G3 10 GB is once again at the back of the field, with the Kingston SSD Now V128 GB (SNV425) just a little ahead of it. The Crucial C300 128 GB is the fastest overall, followed by the X25-M 160 GB and the Vertex 2 100 GB.


Page 13
Capacity and energy consumption

Capacity
As you may already know, in contrast to Mac OS X and Linux, Windows still counts KB, MB, GB and TB as 1 KB = 1024 bytes (base 2). Since 1998 however the standard has been 1 KB = 1000 bytes, and 1 KiB (kibibyte) = 1024 bytes. Magnetic storage devices were already using the current standard before 1998, for obvious reasons of commercial advantage. This is also true for SSDs, or almost:


For most models, true capacity is equivalent to the capacity announced. The capacity in GiBs (displayed as GBs in Windows) is lower. OCZ however names its OCZ Summit and the first Vertexes a little strangely, with the capacity announced in GBs actually corresponding to the GiB capacity (give or take 0.6%). The 60 and 120 GB versions of these ranges offer the same capacity as the 64 and 128 GB versions of other manufacturers.

In contrast, the Sandisk G3 120 GB is limited to just that and the same goes for the SandForces whether in their extended versions (60/120/240 GB) or not (50/100/200 GB
Energy consumption
Here are our energy consumption readings, at idle and in load in IOMeter for sequential writes:


In idle, SSDs are very economical as they donít have to power any platters as hard drives do. Note however there are some differences between them, with the lowest consumption SSD at 0.2 watts, against 1.2 watts for the Kingston based on the JMicron JMF618! In load, the Kingston SNV425 also consumes most, with no less than 4.7 watts: the thermal padís starting to make sense. Still with Kingston, the SNPV325, based on the Toshiba controller, is at 3.8 watts. The other SSDs are more economical, with consumption between 1.5 and 3 watts


Page 14
Conclusion

Conclusion
The best SSD controllers of 2009, the Indilinx Barefoot and Intel PC29AS21BA0, are far from eclipsed by the new arrivals. Just the SandForce SF-1200 and Marvell 88SS9174 manage to equal or outdo them in certain areas, without suffering from too many faults.

In contrast the SanDisk SC4, used on the G3, suffers from disappointing random read performance and the JMicron JMF618 and Toshiba T6UG1XBG used by Kingston give poor random read figures. To these drawbacks, we might add the lack of any obvious TRIM impact, in spite of their specs. Of course, these SSDs do offer an experience that is superior to a standard hard drive overall, with a guarantee of 10 years from SanDisk and aggressive pricing from Kingston for the SSDNOW V and pretty full kitsÖ but you can get much better controllers so why make do?


The SandForce SF-1200, here represented by the Vertex 2 100, Corsair F120 and G.Skill Phoenix Pro 120GB comes into its own with the DuraWrite technology. By compressing, on the fly, data written physically to the NAND, it allows you to economise wear to a maximum and, in desktop use, write amplification (ratio between the volume of bytes of flash written and the volume of bytes to be written requested by the system) lower than 1, which is a first!

In terms of peformance, the SandForce impresses with its random write performance: we got 26,275 IOPS with compressed data and 49633 IOPS with non-compressed data. Here thereís a particularity with the Vertex 2 which has a specific SandForce firmware which blows the 10,000 IOPS (40 MB/s for random writes of 4 KB) software limit used on the SF-1200 but not on the SF-1500 that has been designed for the enterprise market. In practice, donít forget that such performance levels are however of no use for desktop use where most accesses are reads.


The Marvell 88SS9174 used in the Crucial C300, is the other competitor for the Indilinx and Intel controllers. The first controller with SATA 6 Gbits support, gives sequential reads of 338 MB/s in its 128 GB version Ė speeds unequaled up till now. This isnít the only thing in favour of the Marvell. Overall, the C300 is the fastest SSD currently on the market, with very impressive random read and write speeds.

Does this mean the Intel and Indilinx controllers have been sidelined? Not really. While the new arrivals are faster when it comes to random writes, they remain at a similar level for reads, which are most common accesses when it comes to desktop usage. Our tests moreover confirm this state of affairs and while the C300 is best for load times, itís closely followed by the Intels, Indilinx and SF-1200. These two old boys are also alone in allowing TRIM support, though not in real time, in XP and Vista.

When it comes to pricing, itís difficult to separate them as theyíre all so closely priced, especially since the arrival of the 60/120/240 GB SandForce versions. Depending on capacity, model and the retailer, pricing generally goes from Ä2.5 to Ä3 per gigabyte.

How will SSDs develop in the future? Weíve been following developments carefully over the last couple of years and it seems clear that there wonít be any great jumps forward in terms of practical performance. The many new SATA 6 Gbits controllers planned for the end of the year should of course be faster, but not necessarily in the most visible areas to the user. The weak link that was the hard drive in our machines has already been perfectly replaced by current SSDs!


The major sticking point is still pricing and here the Intel / Micron partnership seems to have stolen a march on the rest with the announcement of mass production of 25nm MLC NAND. Intel looks set to start using it on the introduction of its new SSDs at the start of the last quarter, with, we hope, a concomittant reduction in price per GB.

With the size of pages and blocks of NAND increasing on these new chips, the controllers that use them will be aiming to make up for the losses in speed for smaller accesses that are inherent in these changes. Moreover, while this flash memory doesn't seem affected and still has an endurance of 5000 write cycles, the reduction in the thickness of the engraving is also likely to reduce the flash lifespan. The algorithms that aim to reduce flash wear to a minimum, such as SandForceís DuraWrite, will then truly come into their own.


Copyright © 1997-2014 BeHardware. All rights reserved.