SSD 2012 roundup: Sixteen 120 and 128 GB SATA 6G SSDs - BeHardware
>> Storage >> SSD

Written by Marc Prieur

Published on November 15, 2013


Page 1


At last! After many long weeks of work, hereís our new SSD report. For 2012, we have decided to focus on the flagship 120 and 128 GB models and limit our selection to SSDs with an SATA 6 Gbps/s interface. We have tested no fewer than seventeen models:

- Corsair Force 3
- Corsair Force GT
- Corsair Perf Pro
- Crucial C300
- Crucial M4
- Intel 330
- Intel 510
- Intel 520
- Kingston V200
- OCZ Octane
- OCZ Petrol
- OCZ Vertex 3 MI
- OCZ Vertex 4
- Plextor M3
- Plextor M3P
- Samsung 830
- Sandisk Extreme

We're also going to take the opportunity of going over numerous aspects of SSD technology. We hope you find the report of interest!
What is an SSD in fact?
First of all, and for those coming out of hibernation, hereís a little recap. An SSD (Solid State Drive) is a storage device made up of flash memory, as opposed to the magnetic platters used on standard hard drives (HDD).

There are numerous advantages to using SSDs over standard hard drives, the first being performance. In addition SSDs don't make the noise HDDs make and are resistant to knocks. Using them is transparent as far as the system goes, as theyíre addressed like hard drives by the SATA controller.

The disadvantage with them comes from the fact that flash memory has a limited lifespan. MLC flash NAND can for example only be written to 3000 to 5000 times when engraved at 24-27nm or 10,000 times at 34-35nm. Fortunately this is compensated for by wear levelling algorithms that share the wear between the cells and completely resolve the issue of flash NAND lifespan, except if youíre rewriting your entire SSD every day, which isnít of course standard usage.

The chips are also limited in terms of data retention, with a new cell only capable of storing data for ten years. This drops to one year at the end of the life of the drive. In practice and now that the technology has been around for four years, we can confirm that the confidence placed by manufacturers in the reliability of flash memory is justified.
Whatís new?
Whatís new since last yearís review? After all the early activity, the world of SSDs seems to have calmed down a great deal and the flagship models of last year, such as the Crucial M4 and the OCZ Vertex 3, are still around.

Of course new models have also been released but, as we're already at the limit of what the SATA 6 Gbp/s interface will allow in terms of reads, they stand out mainly due to their higher write speeds, notably the smaller capacity models (Samsung 830 and Plextor M3P for example). Other manufacturers have focussed more on random accesses, reaching new records in this domain (OCZ Vertex 4).

Recent months have also shown us that no one is immune from firmware issues either. In August, Intel released a firmware for its 320 SSDs that aims to correct a bug that sometimes made the SSD unusable in the case of an unexpected power outage. In October SandForce brought out a new firmware to end the blue screen issues some users had when coming out of standby. In January, Crucial updated its M4 to deal with a bug that was causing a blue screen every hour after the SSD had been used for 5184 hours and Samsung put a firmware online for the Samsung 830, correcting an issue with a blue screen when coming out of standby. Note however that while this list may appear slightly alarming at first sight, we should say that these problems were not systematic. Thankfully, the great majority of users enjoy the use of their SSDs without coming across these problems!

Page 2
Optimising your SSD

Optimising your SSD in Windows
There are numerous guides on the Internet offering to help you to optimise your SSD in Windows. The first thing to do, which has nothing to do with the OS however nor even with SSDs, as it also applies to standard hard drives, is to activate the AHCI mode in the Serial ATA controller instead of the standard IDE mode, which isnít necessarily the default setting. AHCI mode allows the storage peripheral to optimise processing during multiple simultaneous accesses. Note, itís preferable to make this change before installing your OS as otherwise you'll get a blue screen when you start your computer up again, though Microsoft does offer a "Fix It" for Windows Vista or 7 which allows you to move to AHCI without having to reinstall.

In terms of Windows 7 optimisations, we have to keep in mind that Windows was designed with native SSD support and thereís therefore no need to carry out the slightest performance optimisation at the level of the OS itself.

Behind the optimisation advice claims lie a total under-playing of the robustness of current SSDs, which have been designed to handle a minimum of 20 to 40 GB of writes per day for five years, which is much more than the 5 to 10 GB that youíd expect from standard use. Thereís therefore little point in trying to minimise writes and it can even be counter productive for the overall speed of the system if you move certain temporary files to a hard drive. Equally, indexing your searches is less of an advantage on SSDs than HDDs, though indexing your searches will make them faster than standard searches on SSDs.

This isnít however the case in Windows Vista where you will need to turn automatic defragmentation off as this is counter-productive on SSDs (the OS isn't aware of the internal organisation of data on the flash). Nor does Vista support the TRIM command in real time, unlike 7: with Vista therefore itís best to go for SSDs with software that allows you to run TRIM regularly on the free space on the drive, as with the Intel Toolbox and Samsung SSD Magician.

If youíre using Windows XP youíll also have to create the partitions manually (or with another more recent OS) with SSD-friendly alignment: the basic alignement creates the first partition as of the 63rd sector of a storage device (namely at 32.5 KB), which falls between two flash pages on an SSD. A single 4 KB write would then fall between two 4 KB pages on the SSD and it would therefore have to write both these pages. Windows Vista and 7 leave a margin of 2048 sectors, which helps the SSD to avoid unnecessary wear and poor performance.
Max OS X and Linux

What about other operating systems? Appleís Mac OS X supports SSDs correctly but TRIM support, which has been included since version 10.6.8, doesnít work unless the SSD is from Apple. Trim Enabler however provides a get-around for this. Linux has supported TRIM since version 2.6.33 of the kernel in ext4, but to turn TRIM support on you have to make the partitions with the discard option.
Increasing available capacity in Windows
While Windows 7 doesnít need to be optimised in terms of performance, it can however be modified to free up space on your SSD if youíre using it as your system disk. Two features can in effect take up a relatively significant amount of disk space with respect to the still fairly reduced size of SSDs:

- Hibernation
- Virtual memory

Hibernation reserves space on the disk according to the size of your RAM via the hiberfil.sys file and disabling it saves you the space taken up by this heavy file at the root of the SSD system. This does however mean that you can no longer use the ĎSuspend To Diskí type standby.

To delete hiberfil.sys, you have to execute the powercfg -h off command

The virtual memory is a memory zone on the system disk that can be used like the RAM if the RAM isnít available in large enough quantities. Of course, if you do have to use it, virtual memory is much slower than the RAM but itís availability does mean applications don't crash simply due to lack of memory. Once again here, Windows automatically reserves the equivalent of your RAM on the SSD via the pagefile.sys, and itís possible to reduce the amount of space reserved.

To reduce pagefile.sys you have to customise the size of the Pagefile by going into advanced system settings then
clicking the settings button and the advanced tab and then the modify button for the virtual memory (phew!).

Itís impossible to give a magic formula however (in spite of what many people may say), as everything depends on how you use your PC. If you reduce your Pagefile too much and a software needs more memory than the system can offer, it will crash. Given the low cost of memory however, PCs are generally overprovisioned in RAM compared to their real requirements.

With a general usage level of 4 GB of RAM, you shouldnít need much more than 2 GB of swap and with 8 GB, 1 GB should cover most Pagefile needs.

On a 64 GB SSD with 16 GB of RAM Ė which it is true, represents an extreme case Ė prolonged standby and the Pagefile take up 28 GB of space by default!

Page 3
Controllers and flash

Controllers: Indilinx, JMicron, Marvell, Samsung and SandForce
At the heart of an SSD are the controller and flash memory. On the SATA 6G SSDs in this comparative, various controller chips are used:

- Indilinx Everest: OCZ Petrol et Octane
- Indilinx Everest 2.0: OCZ Vertex 4
- SandForce SF-2281: Corsair Force 3 and Force GT, Intel SSD 520 and 330, OCZ Vertex 3 Max IOPS, Sandisk Extreme
- Marvell 88SS9174: Corsair Performance Pro, Crucial C300 et M4, Intel 510, Plextor M3 and M3P
- Samsung S4LJ204X01: Samsung 830
- JMicron JMF661: Kingston V200

From a purely hardware point of view, the Indilinx chips are in fact (again!) Marvell chips, the 88SS9174 for the Everest and the 88SS9187 for the Everest 2.0. The clocks are however higher and the firmware has been developed by Indilinx. As the firmware is more important than the hardware when it comes to controllers, OCZís choice, though still debatable, is to some extent justified.

Firmware considerations also hold when it comes to the SandForce controller which seems to leave little room for manoeuvre with respect to firmware modifications (the Corsair, Kingston, OCZ and Sandisk implementations are all comparable), with just Intel apparently truly modifying it. The Marvell firmware has been extensively customised by Crucial, Intel and Plextor though there has been no rebranding on the chips.

From a hardware point of view, all these controllers use one or more ARM type processors for processing the algorithms managing their flash. They use an SATA 6G interface to communicate with the host machine and manage the flash across eight channels (except the JMF661 on four channels), with asynchronous or synchronous buses and, with the exception of the JMicron, can use AES data encryption.

While all these controllers have a small amount of memory designed to store the firmware and a minimum of user data (128 KB and 64 KB respectively on the Everest for example), they also all call on external DRAM, mainly as a write cache (up to 1 GB on the Everest 2.0). All except one, the SandForce SF-2281, which makes do with its internal cache.

With respect to flash wear, they all support wear levelling, an algorithm thatís designed to spread wear across the flash cells at the heart of the SSD, as well as more or less extensive ECC error correction to ensure data integrity.

The SandForce SF-2281 is the first SSD to have introduced a technology known as RAISE, which acts as a sort of internal RAID 5 array. Data and parity bits are spread across the SSD's NAND dies to protect it against the loss of data should some of the flash chip be lost. Of course, this technology does have an impact on performance and on available capacity, taking up 8 GB on those SSDs that use it. SandForce has since been followed by Indilinx with a similar technology, RNA, on the Everest 2.0. Marvell have included a comparable feature on the very recent Marvell 88SS9187 (a clue to the chip used in the Everest 2.0!).

Another SandForce technology, which however hasnít been copied, is real time compression of data Ė this is something weíll come back to further on.
Flash: Flash Forward (Toshiba and Sandisk), Hynix, IMFT (Intel and Micron) and Samsung

The other essential part of an SSD is of course its flash memory. Numerous combinations are on offer among the SSDs in this test and we have tried to avoid repetitions as much as possible so as not just to catalogue a series of clones:

The same controller can be used with several types of NAND. The first SATA 6G SSDs used 32 or 34nm memory made by IMFT (a joint Intel and Micron project) or Flash Forward (Toshiba and Sandisk) but weíre now onto 24 or 25nm memory. The OCZ Petrol is the only SSD to use Hynix memory and all we know is that it's engraved at 2xnm (ie between 20 and 29nm). The Samsung 830 uses 27nm Samsung.

Note that there are two types of 25nm IMFT flash MLC, one certified for 3000 write cycles and the other for 5000: only the Intel SSD 520 and the Kingston HyperX use the 5000 cycle flash. In practice however, the wear tests on some SSDs, such as the Crucial M4, have shown that the 3000 flash is capable of going far further and that in any case 3000 write cycles is enough for standard usage.

The choice of communication bus used between the flash and the controller is also important: it can be a so-called asynchronous ONFI 1.0 type bus that is limited to 50 MT/s or an ONFI 2.1 type synchronous one that can then extend to 200 MT/s. Toshiba, Sandisk and Samsung use a similar type of bus to the synchronous ONFI known as Toggle. The flash memory using this type of bus is faster but an asynchronous bus is cheaper.

Page 4
SandForce and compression: take note!

SandForce and compression: take note!
Sandorce controllers are alone in using a real time compression algorithm. It helps to minimise the impact of certain writes on the flash, with SandForce highlighting the example of Windows Vista and Office installations where 25 GB of writes to the SSD are reduced to 11 GB with their algorithm.

There's a double advantage here as not only do you wear less flash, which however as we have said isnít yet a problem given the current endurance of chips, but you also have more unused cells on your SSD, which isn't unimportant as you'll be able to see on the next page.

The Corsair Force 3 speed specifications are misleading

Note however, the marketing speak on this technology is very misleading. Most benchmarks for drives only write series of 0s and 1s, which allows the algorithm to function optimally and give compression rates in the order of 7 to 1, while under more normal circumstances when all goes well, as in SandForceís example, weíre closer to 2 to 1 and with files that have already been compressed to some extent (JPEG, MPEG3 or MPEG4 files for example), the algorithm simply doesn't have any impact.

The same goes for the Kingston V+200

This doesnít seem to have a bearing on SSD manufacturers who highlight the scores obtained in unrealistic cases where compression works to a maximum. For example, SSDs such as the Corsair Force 3 60 GB are announced with read speeds of 540 MB/s and writes of 490 MB/s. The only problem is that, in practice, speeds are a long way down on this and OCZ, who are marketing a similar model, the Agility 3, announces speeds of 525/475 MB/s but also reads of 180 MB/s and writes of 65 MB/s when the algorithm isnít functioning. These more realistic figures are obviously less impressive and arenít highlighted as often, so donít necessarily believe what you read in the tech specs!

OCZ is more honnest, but you have to look at the datasheet as the initial product file doesnít mention the difference!

Itís quite easy to calculate the real compression rate you get with a SandForce chip as the SMART information gives you access to the write volume requested by the OS and that actually written to the Flash. From this you can calculate what is known as the write amplification, namely the ratio between what is written to the NAND and what is requested by the host system, a figure that all controllers bring down as close to 1 as possible. SandForce is alone in taking it below 1 in certain cases with its compression algorithm:

- Incompressible sequential write bench: 1.09
- Compressible sequential write bench: 0.15
- Copy of a Windows 7 + 3d Studio Max 2011 + Visual Studio 2011 + Bibble 5 Pro + BattleField 3 image: 0.76

With incompressible sequential writes, the write amplification is slightly above 1, which is perfectly normal as compression doesn't work in this case. With a compressible bench, a factor of 0.15 is obtained (1.5 GB written to the flash for 10 GB requested by the host system), which is completely unrealistic! When we copy our test system image, the 42 GB only occupy 32 GB, an economy of 10 GB, which isnít negligeable. Note however that of these 42 GB, 1 GB is in fact reserved by pagefile.sys and this 1 GB is empty and therefore highly compressible. The image also includes 1.55 GB for the source code for Visual Studio 2011 and 1.25 GB of JPEG photos for Bibble 5 Pro.

Between the compressible and incompressible readings then, thereís a big difference both between reads and writes on some SSDs. Letís see what happens on some SSDs when we read the data. There are three distinct zones on these screenshots:

- A corresponds to the system image previously described
- B corresponds to a very compressible 4 GB file then an incompressible 4 GB file created via IOMeter
- C corresponds to the unused area of the SSD

[ Corsair Force 3 ]  [ Corsair Force GT ]  [ Crucial M4 ]

On the unused part of the Corsair Force 3, a SandForce SSD with asynchronous memory, weíre at the theoretical maximum of close to 510 MB/s. But in practice on the partition with data weíre far below this at just 272.5 MB/s on average, with a minimum of 187 MB/s and a maximum of 509 MB/s, which corresponds among other things to the zone occupied by pagefile.sys. On the compressible test file weíre at around 425 MB/s and 190 MB/s on the incompressible one.

On a Corsair Force GT, this time with synchronous memory, the impact is much lower for reads as the average on the partition with data is 445.2 MB/s, with variations between 411 and 512 MB/s. Finally on the Crucial M4, which doesnít use a compression algorithm, speeds barely vary at all: 490.6 MB/s on average on the partition with data, as against a maximum of 511.8 MB/s on the unused area.

To avoid becoming victim to the marketing speak of the SSD manufacturers using a SandForce controller, all the synthetic performance tests on the SSDs in this review will be carried out with imcompressible data.

Page 5
A traduire

A traduire

Page 6
Review: Corsair Performance Pro

Review: Corsair Performance Pro
The Corsair Performance Pro is the high end Corsair SATA 6G SSD. Like the Plextor M2P (which is apparently produced in the same factory), it uses the Marvell 88SS9174 controller and Toshiba Toggle NAND engraved at 32nm.

9.5mm thick, this 2.5" SSD comes with a 3.5" adaptor and screws in its box version. Available in 128 and 256 GB versions, it comes with a 3-year manufacturer guarantee.

As of the 128 GB version the official write performance levels announced are very good.

Inside we can see that the Marvell controller comes with two 256 MB DDR3-1133 Nanya chips which serve as the cache. Eight 16 GB Toshiba chips are used for the NAND.

Page 7
Review: Crucial M4 and C300

Review: Crucial M4 and C300
Crucial, a subsidiary of Micron, was the first to use the Marvell 88SS9174 controller in its SSDs. The Crucial C300 was thus the first SATA 6 G SSD to be released, in February 2010. Using 34nm IMFT (an Intel and Micron joint-venture), the C300 was joined by the M4 at the beginning of 2011 with IMFT memory engraved at 25nm and a synchronous bus and improved sequential access performance. As of the 128 GB version the official write performance levels announced are very good.

Note that these SSDs are in the 2.5" 9.5mm format but that 7mm versions are also available. The 9.5mm version can however easily be transformed into a 7mm as the additional 2.5mm are down to a movable intermediary shell, though removing this does invalidate the 3-year guarantee.

The basic Crucial M4 doesnít come with any accessories but a version with a transfer kit (cloning software and USB connectivity) is also on sale to facilitate migration from a hard drive. The Crucial M4 is available in 64, 128, 256 and 512 GB versions.

The official C300 and M4 specs show that the M4 has improved sequential read, sequential write and random write speeds but it performs less well when it comes to random reads. The write performance of the M4 SSDs goes up with capacity up to 250 GB but not any further.

Looking inside the Crucial C300 thereís a Marvell 88SS9174 controller with a 128 MB Micron DRAM cache and sixteen 8 GB 34nm IMFT chips.

The Crucial M4 also uses a Marvell 88SS9174 with 256 MB of Micron DRAM and eight IMFT 25nm 16 GB chips.

Page 8
Review: Intel SSD 520, 330 and 510

Review: Intel SSD 520, SSD 330 and SSD 510
After launching SSDs with its own SATA 3G controllers in 2008 with the X25-M, Intel decided not to develop a new generation controller for SATA 6G and instead use third-party chips, focusing on its firmware instead.

The first SSD based on this new strategy was the Intel 510 based on a Marvell 88SS9174 controller and 34nm IMFT memory, similar to the combination used on the C300 but quite different in practice because of the respective firmware offerings: the Intel 510 has higher sequential speeds to the detriment of random accesses and a default over-provisioning of 8 GB on the 120 GB version and 6 GB on the 250 GB version.

Intel then launched the Intel SSD 520 based on the SandForce SF-2281 and 25nm synchronous IMFT flash certified for 5000 write cycles, as is also the case with the Kingston HyperX. The Corsair Force GT, OCZ Vertex 3 and HyperX 3K all use memory certified for 3000 cycles. Here again the difference comes in the firmware. Available in 60, 120, 240 and 480 GB versions, the 120 GB version also stands out from other SandForce SSDs of this size as Intel doesn't reserve 8 GB for RAISE technology, instead reserving 8 GB for over-provisioning. According to Intel the RAISE feature isnít of any use with the quality of NAND currently used in these SSDs.

Finally, the Intel SSD 330 is a model with slightly lower performance than the 520, and has a 3-year guarantee, like the 510, rather than 5 years on the 520 (within the write wear limits measured by the SMART value of the wear out media for the OEM version).

They can be bought alone or with a 3.5" adaptor plus screws and an SATA cable and Molex to SATA power adaptor. The SSD 510 and 330 are 9.5mm thick like the SSD 520, but the SSD 520 can easily be reduced to 7mm by removing a black plastic overlay from the top of the SSD, though this does annul the guarantee. Note that Intel offers a pretty thorough piece of management software in Windows, which among other things allows you to launch TRIM on the free space in Windows XP and Vista.

The official spec for the SSD 510 claims quite modest random access performance, both for reads and writes. Sequential speeds are however better.

The SSD 520 performs quite differently with different types of data, with an increase in sequential writes on the higher capacity models and a dip in random write speed on the 480 GB version. The spec of the 330 doesnít detail speeds for incompressible data.

Inside the SSD 510 120 GB there are sixteen 25nm IMFT chips which are addressed by a Marvell 88SS9174 chip with 128 MB of Hynix DRAM for cache support.

The SSD 520 120 GB uses a SandForce SF-2281 controller and eight 25nm IMFT chips.

Except for the colour of the PCB, thereís no apparent difference between an SSD 330 and an SSD 520.

Page 9
Kingston V200 en test

Review: Kingston V200
A worldwide giant in memory manufacturing, Kingston has been making moves on the SSD market for quite some time now with its SSDNow range. Initially, Kingston simply resold rebranded Intel X25-Ms. Then the V and V+ ranges were made in association with Toshiba, which supplied the NAND for both and an in-house controller for the V+, with the V using a JMicron chip.

Since then, Kingston has moved closer to more standard solutions and uses the SandForce SF-2281 with synchronous 25nm IMFT for the HyperX (the same as the Corsair Force GT and OCZ Vertex 3) and the SandForce SF-2281 with asynchronous 25nm IMFT for the SSDNow V+ 200 (equivalent to the Corsair Force 3 and OCZ Agility 3). So as to test as wide a variety of solutions as possible we tested the Kingston SSDNow V200 as it is one of the rare SSDs to use the JMicron SATA 6G controller, the JMF66x.

The SSDNow V200 is 7mm high in its 64 and 128 GB versions and 9.5mm in its 256 GB version. It's guaranteed for three years and is sold both alone and in a box version as:

- PC laptop update kit: Cloning software, an installation video, an overlay to take the height of the SSD to 9.5mm, a USB 2.0 casing
- PC desktop update kit: Cloning software, installation video, a 3.5" adaptor, an SATA cable and a Molex to SATA power supply adaptor
- full kit: Everything!

We can only salute Kingston for its efforts on these different box versions.

The official specifications show that the SATA 6G interface isnít saturated, even in reads. Random writes also seem to be a weak point on this SSD, although the level reached is more than enough for desktop usage.

Inside, there are eight Toshiba 24nm MLC flash NAND chips though apparently no Toggle type bus. They're accompanied by 2x128 MB of ProMos DDR2 with a Toshiba controller, which is in fact a JMicron SATA 6G controller. In view of the performance levels announced, it should be the JMF661 which manages flash on four channels, compared to eight on the JMF662.

Page 10
Review: OCZ Vertex 3 Max IOPS

Review: OCZ Vertex 3 Max IOPS
We havenít included the OCZ Vertex 3 and Agility 3 in this comparative as theyíre similar to the Corsair Force GT and Force 3. OCZ does however offer an original SSD that uses the SandForce SF-2281, the Vertex 3 Max IOPS.

This SSD in fact combines the SandForce chip with faster and more durable memory than the synchronous 25nm IMFT, namely 32nm Toshiba Toggle. Otherwise itís standard with 120 and 240 GB models available, with 8 GB reserved for the RAISE on one and 8 GB for the RAISE + 8 GB for over-provisioning on the other.

The Vertex 3 MAX IOPS is 9.5mm high, comes with a 3.5" adaptor and its screws and is guaranteed for three years.

Like Intel, but unlike Corsair or Sandisk, OCZ gives performance for both compressible and incompressible data, which is a good thing. Note that the difference in favour of incompressible data when it comes to random reads is linked to the fact that the test wasnít carried out with the same software or on the same size test file.

Inside we can see the SandForce SF-2281 and eight 32nm 16 GB Toshiba Toggle chips. There's no DRAM type cache, as is the norm with SandForce.

Page 11
Review: OCZ Vertex 4, OCZ Octane and OCZ Petrol

Review: OCZ Vertex 4, OCZ Octane and OCZ Petrol
Launched at the end of 2011, the OCZ Octane and Petrol models both have an Indinlinx Everest SATA 6G controller. Of course, OCZ bought Indilinx last year, allowing it to differentiate itself from SSDs using SandForce controllers. The OCZ Vertex 4 announced at the beginning of April 2012 uses version 2 of the Everest. Note that if the Everests are exclusive to Indilinx from a firmware point of view, the hardware that executes it is none other than a Marvell controller.

The combinations are as follows:

- Petrol: Indilinx Everest + 2xnm Hynix flash
- Octane: Indilinx Everest + synchronous 25nm IMFT
- Vertex 4: Indilinx Everest + synchronous 25nm IMFT flash

They're all 9.5mm thick and only the Vertex 4 comes with a 3.5" adaptor plus screws. Thereís a 3-year guarantee for the Petrols (64 and 128 GB) and Octanes (64, 128, 256, 512 GB and 1 TB) and a 5-year guarantee for the Vertex 4s (128, 256 and 512 GB).

The official specs for the OCZ Petrols indicate a fairly modest performance for random reads, with the Octane doing a bit better. The Vertex 4 stands out most here with very high random performance, with sequential write speeds very much dependent on capacity.

Inside the Petrol we can see an Indilinx IDX300 controller, two 256 MB DRAM chips and sixteen 2xnm 8 GB Hynix flash chips, in principle without a high performance Toggle type bus or synchronous ONFI.

Inside the Octane there is of course an Indilinx IDX300 controller and two 256 MB DRAM chips as well as sixteen synchronous 8GB 25nm IMFT flash chips.

Finally, the Vertex 4 is equipped with an Indilinx IDX400, two 512 MB DRAM chips and sixteen synchronous 8 GB 25nm IMFT flash chips.

Page 12
Review: Plextor M3 and Plextor M3 Pro

Review: Plextor M3 and Plextor M3P
For the older readers Plextor will be synonymous with quality products, notably in the domain of CD players and rewriters. Around three years ago Plextor was bought by PLDS, a joint-venture created in 2007 by Philips and Lite-On IT to make optical disk drives (ODDs), and since then the company has been trying to break into the SSD market with SSDs based on the Marvell 88SS9174.

The Plextor M3 and M3P are the latest Plextor models and are both based on the Marvell 88SSE9174 and 24nm Toshiba Toggle flash. The M3P has a firmware that enables it to obtain even better write performances. The M3 is 9.5mm high and is available in 64, 128 and 256 GB versions while the M3P is 7mm high and comes in 128, 256 and 512 GB versions.

Both come with a 3.5" adaptor with screws, the Acronis True Image HD OEM cloning software and 5-year guarantees. Even better, for the first three years Plextor will pay for a transporter to come and pick up the SSD from you in case of failure (in the European Union, Norway and Switzerland, one year elsewhere). This is much appreciated.

The specs show that the M3P offers a big sequential and random write improvement over the M3, especially on the 128 GB version. The M3P 512 GB model is however slower in terms of random performance than the 256 GB. Note also that the 64 GB version of the M3 doesnít seem to suffer too much in terms of performance.

Inside the Plextor M3 there’s a Marvell 88SS9174 with 256 MB of DRAM cache (this is 128 MB on the 64 GB version and 512 MB on the 256 GB) and eight 16 GB 24nm Toshiba Toggle flash chips.

The Plextor M3P isnít really any different with a Marvell 88SS9174 controller, 256 MB of DRAM cache (512 MB for the 256 and 512 GB versions) and eight 16 GB 24nm Toshiba Toggle flash chips.

Page 13
Review: Samsung 830

Review: Samsung 830
Samsung stands out from the rest in this review as, now that Intel has stopped making controllers, itís the only manufacturer that can claim to master the full SSD production process, from the controller to the flash memory and DRAM memory cache and of course the firmware.

The Samsung 830 therefore uses an in-house controller, the S4LJ204X01, in-house DDR2 memory and in-house 27nm Toggle flash memory. The Samsung 830 is 7mm thick, is available in 64, 128, 256 and 512 GB stand alone versions as well as desktop (3.5" adaptor, SATA cable, Molex to SATA adaptor, Norton Ghost 15) and Notebook (layer to take it to a thickness of 9.5mm, an SATA to USB cable, Norton Ghost 15) kits in a similar offering to Kingstonís.

They're guaranteed for three years and come with a fairly thorough utility that allows you to launch TRIM on the free space in Windows XP and Vista, also possible with the Intel Toolbox.

The official performance levels given are high, except for random reads where other SSDs are a good deal faster. Note that write performance increases as you go up the range from the 64 GB to the 128 GB and 256 GB models, with the graph stabilising with the 512 GB. Even the 64 GB version offers very decent performance.

Inside the Samsung 830 then, there's a Samsung S4LI204X01 controller with 256 MB of Samsung DDR2 and only four 32 GB Samsung flash chips. Of course you canít get this amount of memory on a single die and each chip has actually four dies. This memory is engraved at 27nm and uses Toggle type bus mode.

Page 14
Review: Sandisk Extreme

Review: Sandisk Extreme
After developing its own controllers, namely the G3s and G4s, over the last year Sandisk has been using the SandForce chips among others.

While the Sandisk Ultra uses a SandForce SF-1200 SATA 3G, the Sandisk Extreme uses the usual SandForce SF-2281 SATA 6G. The originality of this SSD comes in its 24nm Toggle flash made in partnership with Toshiba at its Flash Forward joint-venture.

The Sandisk Extreme is 9.5mm high. It's guaranteed for three years and comes in a box that only has a short manual inside it in addition to the SSD. The fact that there isnít a 3.5" adaptor isnít really a problem however, given that all modern casings support 2.5" SSDs natively.

The official specs only give figures for highly compressible data, with therefore very high sequential speeds (writes illusory). As we can see however, random reads get faster as we go up the range and the 480 GB version suffers in comparison to the 240 GB version when it comes to random writes.

Inside the Sandisk Extreme youíll therefore find the SandForce SF-2281 and four 32 GB 24nm Sandisk Toggle flash chips, each with four 8 GB dies.

Page 15
Sequential throughputs

Sequential reads
We started with sequential read speeds measured with IOMeter using 2 MB blocks of incompressible data with 1, 2 and 4 simultaneous commands for two minutes. We give the average obtained in these three cases:

Heading up the field we have several SSDs that can give speeds of over 500 MB/s: Crucial M4, Plextor M3, Plextor M3P and Samsung 830. At 493 MB/s, the Intel SSD 520 scores just below this level, with the 330 slightly less rapid.

Bringing up the rear is the Corsair Force 3, which, in spite of its official spec, suffers due to its asynchronous NAND with reads of just 198 MB/s, hardly worthy of an SATA 6G SSD. Even the old X25-M is faster. While the Kingston V200 and OCZ Petrol do better than the X25-M, their scores arenít fantastic either.
Sequential writes
Still with IOMeter, sequential write speeds were taken on 2 MB blocks of incompressible data and a single command for two minutes:

Here thereís a big difference between the models tested, with just two SSDs scoring above 300 MB/s. These were the Plextor M3P and Corsair Performance Pro while the Samsung 830 manages 299 MB/s. These SSDs are a long way out in front of the competition with the next being the Vertex 3 Max IOPS at 236 MB/s.

In last place is the X25-M V2 but this is understandable given its age. The same goes for the Crucial C300. As with reads, the Corsair Force 3 and OCZ Petrol are among the slowest so-called new generation SSDs, but thereís a smaller gap with the competition than on reads.

Page 16
Random reads

Random reads
Still with IOMeter we took a reading of performance during random reads by blocks of 4 KB on the whole SSD after it had been filled with incompressible data.

These readings were carried out with between 1 and 32 simultaneous accesses over two minutes, which allows us to highlight the ability of SSDs to process these accesses in parallel. However, given that in standard usage the number of simultaneous accesses is most commonly between 1 and 4, we have only given results up to 8 commands to make the graphs more legible.

For information, in addition to random accesses we have also given the values for the same accesses when carried out sequentially.

[Random 4 KB reads: IOPS ]  [ Random 4 KB reads: MB/s]
[Sequential 4 KB reads: IOPS ]  [ Sequential 4 KB reads: MB/s ]

The fastest SSD on random accesses is the OCZ Vertex 4, closely followed by the Crucial C300 and the Plextor M3 and M3P. Overall SandForce SSDs donít do very well here when the exercise is carried out across the whole SSD (they do better on a reduced area), but the OCZ Petrol brings up the rear a good way behind the Octane.

If we look at the same accesses when carried out sequentially, the Samsung 830 dominates with a single command, the Crucial M4 with 2 and 4 and the Plextor M3 and M3P with 8. Two of the SSDs however behave rather strangely: the Kingston V200 isnít any faster with sequential accesses than random and the Vertex 4 is actually slower with sequential accesses! Thereís obviously an issue with the firmware here, given the performance gap to competitor SSDs.

Page 17
Random writes

Random writes
Now we move on to random writes on 4 KB blocks of incompressible data across the whole SSD. The test was carried out for two minutes with between 1 and 32 simultaneous accesses. Here again to make the graphs more legible and given common desktop SSD usage, we have only reported the results on the graph for 1, 2, 4 and 8 accesses. In addition to random writes, we have also given the results for sequential writes.

[ Random 4 KB writes: IOPS ]  [ Random 4 KB writes: MB/s]
[Sequential 4 KB writes: IOPS ]  [ Sequential 4 KB writes: MB/s ]

The Kingston V200ís poor level of performance stands out straight off. In fact, right at the beginning of the test, the SSD manages 5000 IOPS but then drops to 200 IOPS and gives the low two minute average we can see in the graph. You might think that the speed in the DRAM cache would be taken first and then the actual speed, but if you pause after speeds drop down to 200 IOPS and redo the test, it's still at the same level though the cache should have been reset to 0.

Another SSD with rather chaotic performance is the Crucial C300. After managing 17K IOPS for the first minute, it falls to 3K IOPS. This seems to be linked to the latest firmware used as we didn't see this previously. It only happens with a single command. The other SSDs offer relatively stable performance.

With a single command, the SandForce SSDs with synchronous or Toggle memory dominate in spite of the fact that the data is incompressible. The fastest is the Sandisk Extreme. The Samsung 830 and the Crucial M4 aren't far behind. We should also say that with the exception of the Kingston V200, all these SSDs offer more than sufficient performance for desktop use which doesnít require this type of massive write.

With two simultaneous commands the SandForce models suffer and the Crucial M4 takes the lead. With 4 commands however the Corsair Performance Pro performs best and with 8 the Plextor M3P does best.

Putting sequential writes together with random writes highlights the capacities of the controller with respect to write combining, namely concatenating random writes into sequential writes at the level of the flash. The closer random performance is to sequential, the more efficient the controller. With a single command only two SSDs are notably slower in random performance: the X25-M V2 and, even more so, the Kingston V200. Beyond this thereís a dip in overall efficiency except on the following SSDs: Corsair Performance Pro, Crucial C300, Crucial M4, OCZ Vertex 4, Plextor M3 and M3P.

Page 18
Practical tests

Practical tests
Next we carried out the purely practical tests, namely various timed operations after copying the system image on each of the SSDs:

- Booting Windows 7
- Booting 3D Studio Max 2011
- Booting 3D Studio Max 2011 + Visual Studio 2010 + Bibble Pro 5
- Rescan of Ogre source code in Visual Studio 2010
- Regeneration of thumbnails from a directory of 48 RAW files in Bibble 5 Pro
- Launch of Battlefield 3
- Launch of a level in Battlefield 3

Note these timings are not comparable to those of previous tests. Firstly, the processor was overclocked to 4.5 GHz here. Secondly, for Windows 7 we took a reading of the time between the start of the Windows boot (following the boot menu that can be accessed using F8) and the appearance of the desktop.

For 3d Studio Max 2011 we measured the time between launch and the appearance of the tips window, while the multi-application boot was carried out using a batch. The source code for the 3D Ogre engine was used in Visual Studio 2010, while a repertory containing 48 RAW files from a 5D mark II served as the basis for the test in Bibble 5 Pro. In Battlefield 3 we measured the launch time of the game from the validation of the Origin password and the start of the introduction video, and the load of a level between validation of the resumption of the campaign (mission 7 Ė Thunder Run) and the appearance of the image on screen. All these readings were timed by hand five times. The machine was turned off between each reading. We took the average of the three intermediate scores.

For information, with a Hitachi 7K3000 hard drive, this system takes 25.5 seconds for Windows, 40.9s for 3ds and 63.4s for 3ds/VS/Bibble.

We have already noted on many occasions, since our 2010 comparative, that while SSDs do significantly outperform hard drives in applied tests, the differences between SSD models are much slimmer, except when it comes to models that have some kind of a problem. Even the old X25-M Postville does well! The differences you see in the synthetic tests arenít really reproduced in practice for the simple reason that the data is read more rapidly than it can be processed. We showed this a year ago by giving figures that were also very similar when a Ramdisk that was 10 to 15x faster than the SSDs was used.

The significant advantage given by SSDs in comparison to hard drives is obvious here, with just a very slight difference separating the launch of 3d studio max alone and the three application launch. Of course we rarely launch three applications at the same time, but the aim here is to show that SSDs give excellent responsiveness to the system even when the storage system is being used by several simultaneous tasks, a situation in which an HDD struggles.

Itís difficult to designate a winner here as the gaps between different models are quite small, though some so-called new generation models are a bit slower than others. The slowest is the OCZ Petrol, followed by the Kingston V200. Although the OCZ Octane and Vertex 4 are much closer to the rest of the pack, they are also slightly slower. OCZís decision to put everything behind random accesses on the Vertex 4 at the expense of sequential accesses isn't necessarily pertinent.

This time the Hitachi 7K3000 takes 19.2s for Visual Studio and 28.9s for Bibble. Once again, the SSDs all perform very similarly, with just the OCZ Petrol and Kingston V200 a little bit down on the others.

Here the Hitachi 7K3000 takes 58.6s to launch the game and 23s to load a level. Note however that not all games benefit as much from the use of SSDs and given the size of games and the cost per GB of SSDs, it is worth thinking twice before you splash the cash! Once again thereís not much of a difference between the various SSDs and the OCZ Petrol brings up the rear. The Kingston V200 is alongside it when it comes to loading the game.

Page 19
Performance over time & TRIM

Performance over time & TRIM
As weíve already mentioned on several occasions, an SSDís performance levels can deteriorate over time. There are several reasons for this, the first being structural: a hard drive can read, write (unwritten space) or rewrite (space that has previously been written to) data by 4 KB packages. With flash memory you can only read, write or rewrite by 4 KB packages, 4 KB and Ö 512 KB as while you can write or read per page, rewriting must be carried out across a group of pages that make up a block (even 8 KB, 8 KB and 2048 KB for a 25nm 8 GB flash chip and 16 KB, 16 KB and 4096 KB on the forthcoming 16 GB 20nm flash chips ).

When a space thatís already occupied by a file has to be rewritten, there can be an impact on performance and the wear of the flash! Worse still, if the file has been deleted by the OS without TRIM, the SSD doesnít then know that the data that corresponded to this file is now invalid and will carry out a read-modify-write operation instead of simply a write. This problem is however resolved with TRIM, as the OS then tells the SSD that the space is to be considered as unwritten.

This structural issue is accentuated by the existence of optimisations within SSDs aimed at improving random write performance and write amplification. To keep it simple, when a request to write data randomly to an SSD is received, the SSD writes it sequentially to the flash memory, making sure that the addresses in its internal allocation table correspond to the addresses known by the OS (the LBAs) and the corresponding flash pages. For such a mechanism to be efficient, the blocks of flash memory do have to be available, which can be problematic without TRIM support.

What does all this translate to in practice? Testing an SSDs wear isnít easy to do but we have evaluated how the SSDs do when faced with an extreme situation. This time, before taking any performance reading, all accessible space is filled with data written sequentially.

The great majority of this data canít be compressed by the SandForce controller so they don't have a completely artificial advantage, with the exception of 8 GB which is highly compressible and gives some advantage to this technology as will be the case in practice anyway.

A: 80 to 88 GB, fixed incompressible data. B: 8 GB, fixed incompressible data and which will be formatted rapidly for TRIM
at stages 4 and 10. C: 8 GB of fixed highly compressible data. D: Test area previously written with incompressible data.

The only flash cells that the SSD considers as unused are therefore those allocated for over-provisioning and those economised due to compression, something that shouldn't occur naturally in a system that correctly frees up cells using TRIM.

We then took several performance readings at the following points:
1. Filling of SSD
2. 20mn random writes
3. 5mn sequential writes
4. TRIM (freeing up of cells) on 8 GB
5. 5mn sequential writes
6. Reset SSD (secure erase)
7. Filling of SSD
8. 5mn sequential writes
9. 20mn random writes
10. TRIM (freeing up of cells) on 8 GB
11. 20mn of random writes
The 5 and 20 minute periods were fixed according to the speeds and over-provisioning capacities of the 120/128 GB SSDs, notably so as not to write only to the over-provisioning reserve. The TRIM command was only used where indicated, with writes carried out continually otherwise without having recourse to TRIM, which is an extreme case. While we have only given write performance here, we should say that read performance can also be negatively affected, though to a lesser extent. Weíve also included the performance on a Ďnewí (following Secure Erase) SSD, with 100% of the flash then considered as unused by the controller.

[ Corsair Force 3 ]  [ Corsair Force GT ]  [ Corsair Perf. Pro ]
[ Crucial C300 ]  [ Crucial M4 ]
[ Intel X25-M V2 ]  [ Intel SSD 510 ]  [ Intel SSD 520 ]  [ Intel SSD 330 ]
[ Kingston V200 ]
[ OCZ Vertex 3 MI ]  [ OCZ Petrol ]  [ OCZ Octane ]  [ OCZ Vertex 4 ]
[ Plextor M3 ]  [ Plextor M3P ]
[ Samsung 830 ]
[ Sandisk Extreme ]

On this graph you can see a representation of sequential write performance on a "new" SSD, then at stages 8 ("after being filled"), 3 ("after random") and 4 ("after TRIM 8 GB") during a continuous sequential write on the same 24 GB of disk space. Note, this space can be rewritten several times because, at 133 MB/s, this process only takes three minutes.

Only one SSD maintains the same level of performance during sequential writes whatever the situation, the Intel SSD 510. Simply filling it can cause a reduction in performance, notably with the Kingston V200 which then drops from 190 MB/s to around 130 MB/s. The SandForce SSDs also suffer from a reduction in performance here. This isnít immediate and takes between 30 and 60s to kick in because of the additional unused 7 GB of flash available after filling the SSD because of gains due to compression. The Intel SSD 520 and 330 even manage to maintain performance during 90 seconds by adding 8 GB of over-provisioning.

Writing sequentially to an LBA space previously used for random writes and in the absence of flash pages marked as being unused can have a negative impact on numerous SSDs. The majority of SSDs thus fall to a lesser or greater extent under the 100 MB/s bar at the start, with the exceptions being the Intel SSD 510, the Kingston V200 and the Plextor M3 and M3P. With the exception of the 510, they do all suffer from a dip in performance.

Writing sequentially continuously on this same space results in a rapid increase in performance, except on several SSDs: the Corsair Performance Pro, OCZ Octane and Petrol where the gain is small or even inexistent. While freeing up 8 GB of flash via the TRIM command does allow the Performance Pro and most of the SSDs to return to their initial level of performance, the OCZ Octane and Petrol, the Kingston V200 and all the SandForce SSDs are still down: Corsair Force 3 and Force GT, Intel SSD 520 / 330 and Sandisk Extreme.

[ Corsair Force 3 ]  [ Corsair Force GT ]  [ Corsair Perf. Pro ]
[ Crucial C300 ]  [ Crucial M4 ]
[ Intel X25-M V2 ]  [ Intel SSD 510 ]  [ Intel SSD 520 ]  [ Intel SSD 330 ]
[ Kingston V200 ]
[ OCZ Vertex 3 MI ]  [ OCZ Petrol ]  [ OCZ Octane ]  [ OCZ Vertex 4 ]
[ Plextor M3 ]  [ Plextor M3P ]
[ Samsung 830 ]
[ Sandisk Extreme ]

The random write performance on a new SSD is represented on this graph, then at stages 2 ("after filling"), 9 ("after sequential") and 11 ("after TRIM 8 GB") during continuous 4KB writes on the same 24 GB of disk space. Note, this space can be rewritten several times because, at 60 MB/s, less than 7 minutes are required and the test lasts for 20 minutes.

The Kingston V200 shows what its major weak point is here with a level of performance that rapidly drops under 200 IOPS, even on a new SSD. Bizarrely, when the SSD isnít clear of data it performs better (in the order of 900 IOPS)! The Crucial also behaves rather strangely with just a single concurrent access. This is linked to its latest firmware and disappears as of two accesses.

Apart from this, the Vertex 3 does best here. With a very good level of performance on a clean SSD, it manages to maintain it even after being filled with data with sequential writes. Running the TRIM command on 8 GB of the disk and freeing 8 GB of flash brings it back to its initial level of performance and allows it to maintain it even after having rewritten these same 8 GB of flash. The curves arenít as good on the OCZ Petrol and Octane, which suffer quite a bit when the store of flash runs low. Freeing up 8 GB of flash does however allow them to increase performance again.

Remember the SandForce SSDs (Corsair Force 3, Force GT, Intel SSD 520/330, OCZ Vertex 3 MI and Sandisk Extreme) have an advantage here as they have a surplus of available flash resulting from compression, as part of the static data is compressible. In practice however this doesn't give them an enormous advantage compared to the other SSDs, except for the Intel 520 and 330 which have an extra 8 GB for over-provisioning, and only just allows them to maintain their lead over the SSDs which don't have this advantage. Of course, the more the data is compressible, the bigger the reserve of flash theyíll have and the longer theyíll maintain performance at a high level. Strangely, all the SandForces are slower during the first minutes following TRIM.

The Marvell SSDs (Corsair Performance Pro, Crucial M4, Intel 510, Plextor M3, M3P) all perform similarly overall and make gains following TRIM, whether in the first minutes or after stabilisation. The Intel SSD 510 has quite a low level of performance overall and there's a dip after just 14 minutes of writes, which is strange on a 'newí SSD. During this time, the host only requested a little over 40 GB of writes to the SSD and this reduction in performance makes us think that the SSD is short of unused flash, putting the write amplification at x3!

The same phenomenon occurs on the Samsung 830, which does pretty well elsewhere, but with a less significant drop in performance. It begins after around 68 GB of writes, while over the 20 minutes around 92 GB of data is written.

Page 20
Energy consumption and efficiency

Energy consumption and efficiency
Finally we measured energy consumption with a clip ammeter. It was taken at idle and during a sequential read and a sequential write, with this last task being the one that puts most demand on the SSD.

[ Energy consumption ]  [ Energy efficiency ]

Generally speaking, for most of the time the SSD is at idle and this value is therefore important on a laptop or an ultra, where every watt economised extends battery life. The Kingston V200 thus uses a little too much power to be used within a laptop and is then followed by the Everest and Everest 2 SSDs, OCZ's overclocking of the Marvell controller no doubt having an impact here. The most economical SSD is the Samsung 830 with a reading of just 0.4 Watts.

To judge on reads and writes, we looked more particularly at the second graph which shows the speed obtained over Watts consumed. In reads the most efficient is the Intel SSD 510 and the least efficient the Kingston V200. In writes the X25-M is at the back of the field, while the Plextor M3P has by far the best speed to energy consumption ratio.

Page 21

Tieing up a 24-page roundup in a few lines isnít easy, especially as the cost of flash memory is currently dropping, meaning that the price of 120 to 128 GB SSDs is changing almost daily, with some manufacturers responding faster than others. Pricing is however an important criteria on which to judge these solutions.

Above and beyond price, if we had to choose one SSD from all those reviewed here, we would go for the M3 Pro on which Plextor is offering the most successful implementation of the Marvell 88SS9174 controller. Offering performance of the first order in all areas, this SSD also comes with a 5-year guarantee and a 3-year home collection service. Unfortunately, this model is also a good deal more expensive than the restÖ so you are in fact paying your transport costs, on purchase! The Plextor M3, which is a little cheaper, is only down on the M3P when it comes to writes. The Corsair Performance Pro, which is based on the same configuration as the previous Plextor generation, namely the M2P, doesn't perform as well but isn't as affordable.

The OCZ Vertex 4 could have been up there above the Plextor M3P if it had been better balanced. With a 5-year guarantee and more affordable, it has established a new record for random accesses and offers very good random write performance even in the absence of TRIM. Unfortunately, this comes to the detriment of sequential accesses which are particularly low on small sized blocks, a compromise which is useful on some benchmarks or for very specific usage but which in practice doesnít put the Vertex 4 at the top of the field when it comes to standard usage.

If youíre looking for a compromise between performance and pricing, the Samsung 830 seems to stand out from the rest. Always well placed, this SSD also has a very complete utility in SSD Magician, enabling TRIM on free space in Vista and XP (Intel offers the same thing with its Toolbox). Although somewhat lagging in sequential writes, which isn't necessarily an important criteria for a system disk of this capacity, the Crucial M4 is also a very worthy model with very good performance in practice and almost faultless reliability (except for the 5184 hours thing!) according to the user feedback that has been coming out for around a year now since its release. At the end of the day, this is more important than several tenths of a second in the benches!

Though less expensive, the OCZ Petrol and Kingston V200 should be avoided. These SSDs offer lower performance than the competition and we recommend you go for other models from these constructors (OCZ Vertex 3 or 4, HyperX or HyperX 3K from Kingston). Less problematic than the Petrol, the Octane comes midway in the performance table and offers a less sustained performance than most other SSDs during intensive usage.

If youíre absolutely looking for an affordable SSD, we advise you rather to go for a SandForce SF-2281 model. Among these, the Sandisk Extreme does best as its controller and Sandisk Toggle NAND combination allows it to perform at a level close to the versions with synchronous IMFT NAND (the Corsair Force GT but also the OCZ Vertex 3 and Kingston HyperX/HyperX 3K) at the same time as being priced on a level with those with asynchronous IMFT NAND (Corsair Force 3, OCZ Agility 3 ou Kingston V+200)!

Intel also offers a pretty good SandForce model with its Intel SSD 520, in contrast to the Intel SSD 510 which wasnít as impressive. It is more expensive but comes with a 5-year guarantee and a more robust firmware that Intel has reworked, tipping the balance in its favour. Whatís more, if youíre looking for a general consumer SSD to use for managing a database, this is the SSD weíd go for as it uses IMFT NAND certified for 5000 write cycles and this type of compressible load would benefit from the SandForce algorithms, limiting wear and sustaining performance.

If itís too expensive for you, the Intel SSD 330 is worth a look as it's identical to the 520 in terms of hardware, at least in the version tested, with a slight performance limitation in terms of the firmware. It has a 3-year guarantee and while Intel may well eventually use chips certified for 3000 cycles for it, this wonít be an issue at common usage levels.

Page 22

Page 23

Page 24

Page 25

Page 26

Page 27
Test protocol

Test protocol
Testing SSDs correctly requires appropriate software. Tools such as h2bench, HD Tune or HD Tach were originally designed for hard drives and can, for example, operate on zones with no data written to them.

They can then be fooled during read speed readings, as random reads are transformed into sequential reads by some controllers when an SSD is empty! Equally, applications such as ATTO Disk Benchmark only write series of 0s and 1s, which allows controllers with real time compression algorithms to perform better than they would with actual data.

Even specialised software isnít exempt from faults. For example, while CrystalDiskMark allows you to test an SSD quickly, it gives partial results. Its test file is only 4 GB in size in the best case, which means that random tests only address the SSD on a fairly restricted area, which gives the advantage to certain controllers.

We therefore used the formidable box of tricks that is IOMeter to measure synthetic performance. Performance readings were taken in several cases, for two minutes each time:

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

This allowed us to review both storage device speeds and inputs/outputs. The tests on blocks of 4 KB were carried out with 1, 2, 4, 8, 16 and 32 simultaneous commands so as to highlight the controllerís ability to work on accesses in parallel. Given that it's very rare in desktop usage to require more than four simultaneous accesses and so as to give the graphs greater legibility, only the results up to 8 commands are shown.

With sequential reads, we carried out the test with 1, 2 and 4 commands and reported the average of the three. With sequential writes, the test was carried out with just a single command. These tests were carried out with data that cannot be compressed by the SandForce controller real time compression, whereas previously we gave the results with both incompressible and compressible data.

Moving onto the practical tests, we started by looking at write and read speeds for various groups of files. These groups were composed as follows:

- Extra large: 731.17 MB on average
- Large files: 5.2 MB on average
- Medium sized files: 800.88 KB on average
- Small files: 48.78 KB on average

We used an 8GB RamDisk as the source or the target for reads or writes on the SSD. Given the rapidity of recent SSDs and so as to obtain results that are less subject to variation, we used Robocopy with some in-house software that allows us to carry the tests out continuously. We carried out the test five times in a row and give the average of the three intermediate scores.

Next we carried out the purely practical tests, namely various timed operations after installation of Windows 7 64-bit on each of the HDDs:

- Booting Windows 7
- Booting 3D Studio Max 2011
- Booting 3D Studio Max 2011 + Visual Studio 2010 + Bibble Pro 5
- Rescan of Ogre source code in Visual Studio 2010
- Regeneration of thumbnails from a directory of 48 RAW files in Bibble 5 Pro
- Launch of Battlefield 3
- Launch of a level in Battlefield 3

Note these timings are not comparable to those of previous tests. Firstly, the processor was overclocked to 4.5 GHz here. Secondly, for Windows 7 we now take a reading of the time between the start of the Windows boot (following the boot menu that can be accessed using F8) and the appearance of the desktop.

For 3d Studio Max 2011 we measured the time between launch and the appearance of the tips window, while the multi-application boot was carried out using a batch file. The source code for the 3D Ogre engine was used in Visual Studio 2010, while a repertory containing 48 RAW files from a 5D mark II served as the basis for the test in Bibble 5 Pro. In Battlefield 3 we measured the launch time of the game from the validation of the Origin password and the start of the introduction video, and the load of a level between validation of the resumption of the campaign (mission 7 Ė Thunder Run) and the appearance of the image on screen. All these readings were timed by hand five times. The machine was turned off between each reading. We then took the average of the three intermediate scores.

In addition to the 120 to 128 GB SATA 6G SSDs in the comparative, we also included the performance obtained on an Intel X25 M ďPostville" 120 GB, a range launched in July 2009 (see our test)

Finally we carried out tests of sustained write performance that we have detailed on the page dedicated to this subject. The test system was made up of an Intel DP67BG (P67 Express) motherboard along with 16 GB of DDR3-1600 and a Core i7-2600K, all in Windows 7 SP1 64-bit.

Page 28
Reading and writing of files

Reading and writing of files
Moving onto the practical tests, we started by looking at write and read speeds for various groups of files. These groups were composed as follows:

- Extra large: 731.17 MB on average (videos)
- Large files: 5.20 MB on average (audio)
- Medium-sized files: 800.88 KB on average (photos)
- Small files: 48.78 KB on average (various files)

We used a RamDisk as the source or the target for reads or writes on the SSD. Given the rapidity of recent SSDs and so as to obtain results that are less subject to variation, we used Robocopy with some in-house software that allows us to carry the tests out continuously. Here only the small files can be compressed by the SandForce compression algorithm.

[ Reading of files]  [ Writing of files]

The fastest SSD for reads of small files is the Corsair Performance Pro, closely followed by SSDs based on the SandForce and Everest 1 controllers. When the size of files is increased, the gaps between the different models gets bigger and with extra large files the SSDs with the best sequential performance were: Crucial M4, Plextor M3 and M3P and Samsung 830.

The differences are more marked with writes. The Kingston V200 is this time significantly slower than the rest, whatever the type of files used. On small files the SandForce SSDs with synchronous 25nm IMFT memory or 34 nm Toshiba Toggle (Corsair Force GT, Intel 520 et Vertex 3 Max IOPS) are in the lead, with the Sandisk Extreme and Intel 330 a little way behind. With the exception of the V200, the Crucial C300 and M4 bring up the rear.

As the files get bigger the Plextor M3P, Corsair Performance Pro and Samsung 830 take over as the fastest performers. The good old X25-M and C300 suffer here but the OCZ Petrol and Corsair Force 3 arenít so far behind.

Page 29
Capacity, over-provisioning and Garbage collector

Capacity: GBs and GiBs
As with hard drives, the usable capacity of SSDs remains a big mystery for many users in Windows for the simple reason that the OS, SSD manufacturers and NAND manufacturers all calculate capacities differently.

Windows, in contrast to Mac OS X and Linux, continues to count KBs, MBs, GBs and TBs on a basis of 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 sales reasons.

A 128 GB SSD will thus show up in Windows as having a capacity of 119.24 GB, but this is a Windows error that should show 128 GB_or_119.24 GiB. A 120 GB SSD is displayed as having 111.79 GB available instead of 120 GB_or_111.79 GiB.

To make things more complicated, DRAM memory or flash NAND manufacturers still use base 2. Taking a latest generation Micron chip for example, a 128 Gb or 16 GB model is made up of 16 384 blocs, each of which contains 256 pages. Each page is made up of 4096 available bytes and 224 bytes for ECC. This gives us 17179869184 bytes, which is 17.18 GB or 16 GiB and not 16 GB.

All current 120/128 GB SSDs thus have 137.4 GB/128 GiB of flash. They are however sold as having 128 GB/119.24 GiB of available space at best, as the additional available space is reserved for the replacement of defective blocks or the storage of certain pieces of data thatís useful to the controller such as the translation table between the logical block addressing system (LBA) and the physical flash addresses (the pages).

Some SSDs go further, only offering 120 GB/11.79 GiB of usuable space to store your data. The missing 8 GB may be linked to two things on the 120 GB SSDs tested here:

- Redundancy technologies like SandForceís RAISE: Corsair Force 3, Force GT, Sandisk Extreme
- Over-provisioning: Intel SSD 510, Intel SSD 520

Note that on SSDs of different capacities this can change. As RAISE requires 8 GB of flash, on the 60 GB versions of the Force 3 and GT itís deactivated in favour of 4 GB of over-provisioning (OP), while the 240 GB versions of their SSDs, like the Intel 520, have 8 GB of OP and 8 GB for RAISE.
Keeping flash space in reserve (OP) means the controller will always class a significant number of pages and blocks as being free. This enables you to avoid rewriting operations (read-modify-write), which, remember, can only be done per block, when a simple write is done on a per page basis and optimise the use of write combining which concatenates random writes into sequential writes in terms of the flash.

In the absence of available flash, in a worst case scenario to rewrite to an already used block, you have to read it, combine the old and the new data and rewrite it. When we know that on a 32 Gb flash die a page is 4 KB and a block 1 MB, against 8 KB and 2 MB on a 64 Gb die and 16 KB / 4 MB on a 128 Gb die, this can engender significant wear and a dip in performance.

Of course the TRIM command gives us a solution to the problem. To recap, this command is sent to the SSD when a file is deleted and invalidates the data stored on the flash containing it. This gives a relative coherence between the amount of free space as seen by the user and the free flash on the SSD, which isnít the case without TRIM as then all that happens is the file system allocation table is modified. In the absence of TRIM, an SSD without over-provisioning will quickly run short of flash memory that can be considered as free of user data and this limits it in terms of write combining possibilities and means it will often have to resort to read-modify-write.

With some over-provisioning, the SSD will in any case have a volume of flash memory that it can manipulate as it chooses: here it's the physical addresses (pages) that wonít be linked to the logical addresses (LBA) in the SSD translation table. While Intel automatically provides over-provisioning on its 510 and 520 SSDs, itís possible to introduce it onto an SSD yourself: you simply donít use all the space on your SSD, by for example not partitioning it completely. Here SandForce has an advantage that is inherent to its compression algorithm, as even if you use all the user space available on the SSD, the data will occupy less space physically in the flash and part of it will therefore remain available. The volume of available flash will then depend on the compression ratio that the controller manages to obtain.

E9 reports the volume of data actually written to the flash and EA/F1 the volume of data written by the host to the SSD

What is the impact of over-provisioning in practice? To find out more we used a Corsair Force GT as, like all the SandForce SSDs, it allows access, via the SMART, to the write volume requested by the OS and to that actually written on the flash, thus giving us the write amplification.

Here we measured performance during incompressible 4 KB random writes with a single concurrent access for 20 minutes on 24 GB of the SSD without use of the TRIM command in various configurations of available flash space:

- completely empty (new) SSD after secure erase (120 GB of free space)
- SSD with 0 GB of free space and unoccupied flash at the beginning of the test
- SSD with 4 GB of free space and unoccupied flash at the beginning of the test
- SSD with 8 GB of free space and unoccupied flash at the beginning of the test
- SSD with 16 GB of free space and unoccupied flash at the beginning of the test
- SSD with around 0 GB of free space but with 7 GB of unoccupied flash due to the SandForce compression at the beginning of the test

As you can see, the more flash you have available at the beginning of the test, the better the performance of your SSD. This is of course true at the beginning of the test, a phase during which only this free space is used, but performances also stabilise at a higher level after a certain amount of time. You never however return to the initial level of performance which can only be maintained with the use of TRIM.

The reduction in performance is entirely linked to increased pressure on the NAND as the write amplification figures obtained during the test show:

- Empty SSD: 1.39x
- 0 GB of free space: 4.54x
- 4 GB of free space: 2.54x
- 8 GB of free space: 1.96x
- 16 GB of free space: 1.53x

Sequential writes also benefit from this surplus of flash in the absence of TRIM. Here are the performances obtained when the same 24 GB area is written sequentially after it has been written to with random writes:

The amplification factors obtained are as follows:

- Empty SSD: 1.09x
- 0 GB of free space: 1.78x
- 8 GB of free space: 1.24x

The Garbage collector
SSDs use a mechanism called Garbage Collector which aims to reorganize data within the flash so as to maximize the number of blocks that are free of data and enable the following writes to be faster. This Garbage Collector can run in real time when new data is written or during idle time and itís then referred to as Idle Garbage Collector.

Demonstrating the impact of Idle Garbage Collector is quite complex as the SSD has to be put back in the same conditions before beginning each test and then you have to wait the same amount of time between various write phases for idle GC to do its job. On a Corsair Performance Pro, reputed for having an efficient GC, we carried out the following manipulations:

- Fill a partition A of 104 GB and a partition B of 24 GB (0 GB of over-provisioning)
- Fill a partition A of 96 GB and a partition B of 24 GB (8 GB of over-provisioning)
- Random writes by blocks of 4 KB on B (20 minutes)
- Pause of varying amount of time
- Sequential writes by blocks of 2 MB on B (5 minutes)
- Pause of varying amount of time
- Sequential writes by 4 KB blocks on B (5 minutes)

We start with the performance level reached for sequential writes after having written randomly to partition B.

[ 0 GB of OP ]  [ 8 GB of OP ]

With 0 GB of OP, the SSD sees its performance drop a lot on the initial 300 MB/s and starts at under 50 MB/s without a pause between the two types of access. As the sequential writes kick in, performance rises. The idle GC allows you to get a starting speed of 90 MB/s, whether this be after 5 or 30mn of idle time, but the SSD then struggles to get over 100 MB/s. With just 1mn of idle GC the impact is much reduced. Having 8 GB of over-provisioning means thereís less of an impact on performance from the beginning, whether with or without GC. This time, there is a difference between leaving the SSD for 5 or 30mn but the curves for 1 and 5 minutes are the same.

Next we move on to sequential writes after carrying out sequential writes on partition B.

[ 0 GB of OP ]  [ 8 GB of OP ]

With 0 GB of OP and without any pause and therefore without GC, performances start at 30 MB/s as against 75 MB/s on the new SSD. With a pause of 1, 5 or 30mn between the sequential and random writes, performance levels are very similar and you start at 55 MB/s for the first minute but then fall straight back down to the same level, the gain therefore seeming to be linked more to the fact that the write cache has been freed than anything else. With 8 GB of over-provisioning performance levels are better maintained with a constant 50 MB/s. Even with a pause of a minute, we start at around 65 MB/s but after a minute we drop back to the same level we had without any pause.

As you can see, Idle Garbage Collector doesn't make much of a difference. The best way to guarantee a stable level of performance on an SSD is still TRIM and, should it be absent, over-provisioning. Idle Garbage Collector is of secondary importance.

Page 30
Review: Corsair Force 3 and Corsair Force GT

Review: Corsair Force 3 and Corsair Force GT
The Corsair Force 3 and Force GT SSDs make up the heart of the Corsair SATA 6G range. They both use SandForce SF-2281 controllers, with the Force 3 using 25nm asynchronous IMFT NAND and the Force GT using synchronous 25nm IMFT NAND. Of course the combination used on the Force GT is the fastest.

There are numerous equivalents of these SSDs in terms of the controller / NAND combination, the most common being:

- Corsair Force 3: OCZ Agility 3, Kingston V+200
- Corsair Force GT: OCZ Vertex 3, Kingston HyperX/HyperX 3K, Intel 520 and 330

From the outside, these 2.5" SSDs look standard at a thickness of 9.5mm, apart from the colour (original) chosen for the Force GT. They come with an adaptor and screws for a 3.5" slot. With 60, 90, 120, 180, 240 and 480 GB versions, thereís a 3-year guarantee.

The official Corsair Force 3 and Force GT specs only state performance on highly compressible data. Of course here, the scores are marvellous and even the Corsair Force 3 60 GB gives an impressive showing. Unfortunately when you look at the more realistic figures for incompressible data given by its competitor, OCZ, for similar SSDs, they arenít anywhere near as good, dropping from reads of 540 MB/s to 180 MB/s and writes of 490 MB/s to 65 MB/s! With its synchronous memory, the Corsair Force GT suffers less with incompressible data, especially for reads. Performance levels increase with the higher capacity versions, especially read speeds, though there is a dip on the 480 GB versions.

No surprises inside with the SandForce and sixteen 25nm IMFT flash chips offering a total capacity of 137.4 GB or 128 GiB. Only 120 GB or 111.79 GiB are accessible to the user because of the standard reserved space on flash chips (around 7%) which takes the usable capacity to 128 GB or 119.24 GiB with an additional 8 GB reserved for the SandForce RAISE technology which gives protection of your data in the case of total or partial failure of a flash chip. This gives us a usable capacity of 120 GB or 111.79 GB.

No external DRAM is required by the SandForce chip, reducing costs a bit.

Copyright © 1997-2015 BeHardware. All rights reserved.