ADATA GAMMIX S10 in the test

SSDs in the compact M.2 design are becoming increasingly popular in gaming PCs. Since current mainboards are usually equipped with the necessary slots, more and more users want to benefit from the higher transfer rates. For some, however, it still has to be cheap, so today we're looking at a cheap entry-level SSD from ADATA in the M.2 design: the XPG GAMMIX S10.

Intro

SSDs in the compact M.2 design are becoming increasingly popular. This is not necessarily because they take up less space and are therefore easier to install in compact notebooks and computers. Much more decisive is the fact that different logical interfaces can be used on M.2 modules. For example the well-known SATA specifications with speeds of up to 600 MB / s or the much more interesting one for PC enthusiasts PCI Express based on NVM Express. This enables speeds of over 3.200 MB / s, which is a significant leap in performance.

Impressions

Today we're looking at a model from Taiwanese storage manufacturer ADATA: the XPG GAMMIX S10. This SSD was placed in the entry-level area of ​​the ADATA portfolio last fall, and its task is to offer high read rates at a low price and thus outperform conventional SATA models in terms of performance. In this price segment, however, the use of cost-efficient TLC-NAND is inevitable, with which high write rates are usually not to be expected. Most manufacturers usually use SLC caches to compensate for this as much as possible. In this review, we'll see how ADATA addressed this issue.

SSD bookmarks:

• Structure
• Empty or free?
• Wear leveling
• The root of all evil: Read-Modify-Writes garbage disposal
• Spare area and overprovisioning
• TRIM does not delete!
• Speed
• SLC, MLC, eMLC, TLC
• SSD caching - how-to

Recent SSD Reviews:

• Roundup: PCI Express SSDs with NVMe
• Western Digital Blue 500GB SSD
• Crucial MX300 with 1050 GB
• Crucial BX100 with 250 GB
• Corsair Neutron XT with 480 GB
• AMD / OCZ Radeon R7 SSD with 240 GB
• OCZ ARC 100 with 256 GB
• Crucial MX100 with 256 GB
• Corsair Force LX with 256 GB
• Crucial M550 with 256 GB and 1 TB
• SanDisk Extreme II and Ultra Plus
• Samsung 840 Pro and EVO
• SSD reboot

The test candidate

ADATA's GAMMIX S10 relies on the widespread combination of inexpensive 3D NAND with TLC control and an SLC cache to accelerate write access. The TLC control of the memory cells makes it possible to store three bits per cell. The writing process takes longer and the writing rates drop. An SLC cache counteracts this and makes it possible to write a few gigabytes much faster. Of course, the cache is emptied again in the background by moving the data to the TLC area so that the write acceleration is available again later. Apart from the lower production costs, this mixture already more than adequately covers the needs of most home users and gamers: The entire amount of data on the data carrier can be read quickly, while writing processes are usually only necessary in moderate amounts. An overview of the different storage types can be found here .

ADATA calls its cache an "Intelligent Cache", which is usually an indication that the size of the cache is dynamic. Unfortunately, exact information about its size cannot be found in the data sheets, so we will try to derive its size in the test. The Silicon Motion SM2260 controller is a well-known representative. Announced by Silicon Motion in 2015 and used in entry-level SSDs such as the Intel 2017p since 600, it is inexpensive and mature.

Overview of the technical data

In the end customer market, Samsung's 960 EVO and Intel's 600p are direct competitors to our test candidate. The following table compares the technical specifications of the manufacturers again:

Impressions

The most striking feature is the glued-on heat spreader. It is relatively flat and designed with red, dynamic shapes on a black background. In principle, a heat spreader can help to dissipate waste heat from the SSD controller more easily, but it also depends on being supported by the case ventilation - especially in very compact cases.

The 32-layer 3D-NAND can with TLC control 384 Gbit per The save and comes from IM Flash Technologies, a joint venture between Intel and Micron, which also works on high-performance memory 3D XPoint files. We had Intel / Microns 3D-NAND here briefly introduced. As described, the well-known Silicon Motion SM2260 is used as the controller.

The cooler is attached to the circuit board by means of two heat-conducting adhesive pads. However, if you look under the cooler from the side, you can see that most of the controller is not covered by the adhesive pads and thus the heat dissipation from the metal controller surface is at least somewhat limited. Due to the low heat development, this should not be seen as a problem.

The adhesive pads and the cooler itself are 2 mm high in total.

Equipment

ADATA does not provide an NVMe driver for this SSD, so it is addressed with the native driver of the operating system. For all other tasks there is the ADATA toolbox. This supports almost all ADATA models and enables the display of the operating parameters and service life, the optimization of operating system settings with regard to the SSD (e.g. TRIM), firmware updates as well as a quick or full diagnosis of the drive. A read test is carried out over the entire memory area. The package is rounded off by a five-year guarantee.

Test environment

Hardware

Test station:

• CPU: Intel Core i3 3220 - 4 x 3,3 GHz (Turbo: off) [Amazon offers]
• Motherboard: ASUS P8H77M (H77 chipset) [Amazon offers]
• Memory: 8 GB (4 x 2 GB) Team Xtreem - SPD operation: DDR3-1333 9-9-9-24-1T at 1,5 volts [Amazon offers]
• Power adapter: NZXT 650 Watt HALE82 Series [Amazon offers]
• Boot drive: OCZ Vertex-2-SSD as boot drive [Amazon offers]

The test candidate:

• ADATA XPG GAMMIX S10 (Amazon offers), firmware: CB1.1.1

Comparison models:

• AMD / OCZ Radeon R7 240 GB (HT4U test / Amazon offers)
• Corsair Force LX 256GB (HT4U test / Amazon offers)
• Corsair GTX 480GB (HT4U test / Amazon offers)
• Corsair Neutron XT 480GB (HT4U test / Amazon offers)
• Crucial M550 256GB (HT4U test / Amazon offers)
• Crucial M550 1TB (HT4U test / Amazon offers)
• Crucial MX100 (250GB) (HT4U test / Amazon offers)
• Crucial MX300 (1.050GB) (HT4U test / Amazon offers)
• Intel 600p 512GB (HT4U test / Amazon offers)
• OCZ ARC 100 240 GB (HT4U test / Amazon offers)
• Samsung 840 120GB (HT4U test / Amazon offers)
• Samsung 840 EVO 250GB (HT4U test / Amazon offers)
• Samsung 840 Pro 256GB (HT4U test / Amazon offers)
• Samsung 960 Evo 512GB (HT4U test / Amazon offers)
• SanDisk Extreme 240GB (HT4U test / Amazon offers)
• SanDisk Extreme II 240GB (HT4U test / Amazon offers)
• SanDisk UltraPlus 256GB (HT4U test / Amazon offers)
• Toshiba OCZ RD400A 512GB (HT4U test / Amazon offers)
• WD Blue 500GB (HT4U test / Amazon offers)

Software

Our benchmark course

Our benchmark course aims to answer the following questions:

• How fast is the SSD reading and writing large files sequentially and reading and writing small files at random?
• How do fragmented blocks (not to be confused with file fragmentation!) And the resulting read-modify writes affect performance after a heavy write load?
• How fast is the SSD in a continuous load scenario (steady state)?
• Can TRIM restore full performance?
• How effective is garbage collection?
• How fast is the SSD when certain mixes of large and small blocks occur?

Synthetic benchmarks

The use of synthetic benchmarks cannot be avoided, since only with these the technical limits of the SSDs become visible. They show the maximum achievable.

With these benchmarks we determine the performance in the following states:

In this way it can be seen whether and to what extent the performance of the SSD is falling and whether TRIM can restore the original performance.

It does not matter whether you copy a few hundred MP3 or video files or simulate this work with Iometer, the effort is the same for the SSD. Differences resulting from the file system of the operating system then affect all SSDs equally, so that the ratios of the performance differences remain the same.

Trace benchmarks

Real life, on the other hand, can be simulated using trace benchmarks such as PCMark or Iometer profiles, which simulate use cases. With these tests, practical accesses are carried out in a reproducible manner.

For practical results, we carry out these tests after the SSD has already been written with load profiles several times and is occupied with active data except for a remaining 10 GB. This gives you the performance values ​​of an SSD that has already been used and is currently mostly full.

Treatments

We test less per application itself. There are two main reasons for this: First, the CPU limit falsifies the performance gap between the SSDs. For example, when the SSD has to wait for the CPU to process certain data before the SSD can continue working when the application starts. Due to the CPU limit, the SSDs move closer together than would be the case with faster CPUs later. Second, many applications can only be measured with a stopwatch, which is too imprecise for us, especially since the results are sometimes only tenths of a second apart. However, we carry out our long-serving OpenOffice copy test because it is easy to reproduce. We have only increased the amount of data there by a factor of 12. It is now 3,06 GB of data in over 48.000 files of various sizes that will be duplicated on the test drive.

Continuous load measurements

As described in the "Load behavior" section, SSDs break down under continuous random write load if the garbage collection cannot provide free blocks fast enough. Of course, such a load behavior only rarely occurs in normal home use. For some readers, however, it may be interesting to know whether an SSD is also suitable for somewhat tougher use. For example as a data medium for a virtualizer, where many small accesses can occur in parallel, or as a disk for a database test environment.

For this test, we unleash as many 4k writes as possible on the SSD via Iometer and create a graph that shows the performance over time. We repeat this test after a 30-minute or 12-hour break to see whether the garbage collection was able to provide enough free blocks for high performance during this time. Since Iometer works with a large test file, which is never deleted but only overwritten, influences by TRIM in these two repeat runs are excluded. The increase in performance through TRIM itself is then measured in a fourth run. This takes place after a quick format, which "trims" the drive. The test file is then created again.

We would like to point out that this goes well beyond the normal requirements for SSDs for home use. If an SSD does not do so well here, it is therefore not counted negatively. But we want to find out which SSDs stand out from the crowd. In addition, this test makes it easier to see how and whether the garbage collection is working.

MByte / s or IOPS?

Normally, we give the measurement results in megabytes per second. For the profile tests, however, we opted for IOPS (Input / Output Operations per Second = input and output commands per second). An input or output command can mean reading or writing a block. This does not detract from comparability. If a data carrier achieves 128 IO per second in a write test with 1.000 KB blocks, then mathematically this results in 1.000 * 128 KB = 128 MB per second. When an operating system writes MP3 files or videos, it does so in blocks too, and the block sizes ultimately depend on the size of the files and the formatting of the file system. With many small files, this may limit the number of IOPS and with large files the maximum write rate of the SSD. Therefore, it makes sense to use the specification of IOPS wherever there is a high number of read and write operations and / or different block sizes are involved.

In the case of continuous load measurements, the indication in IOPS has the additional advantage that the maximum IOPS information usually advertised by the manufacturers can be compared directly with the real results.

Measurement results

Sequential reading

Because we run the sequential read tests on Iometer with a queue depth of 1 and a transfer size of 2M, not all drives can reach their maximum theoretical read speeds. However, the performance differences with the same queue length are noticeable. AS SSD utilizes the reading process more optimally.

Sequential writing

Since our Iometer test run writes a large amount of data for several minutes, the write rates for this TLC drive are relatively low because the SLC cache is not sufficient for such a large amount of data. It is noticeable that the value (after load) is higher. Intel's 600p behaved in the same way, and both models have the same controller, so that a connection with the way the SLC cache works can be assumed (see next page).

The AS SSD benchmark, on the other hand, writes a smaller amount of data, so it tends to make the higher write rates with SLC cache visible. While users with large amounts of writing (e.g. 4K video editing) should use the Iometer benchmark as a guide, the AS-SSD benchmark is more decisive for most users.

Sequential writing over time

Here we check how the sequential write speed develops over time to put the SLC cache to the test. The controller first writes large amounts of data in an area that is quickly controlled in SLC mode. If this area is full, the data rate drops accordingly. The size of the SLC cache can be derived from the write rate and the point in time at which the write rate dropped. ADATA advertises the cache as “Intelligent Cache”. Some manufacturers combine this with a dynamic adjustment of the cache size, depending on how full the data carrier is. We take the first sample measurement when the SSD is only a quarter full:

The GAMMIX S10 can maintain a write rate of just over 15 MB / s for about 800 seconds before the further write processes take place directly in TLC mode. Now we repeat the measurement if only 10 GB are free on the SSD:

The values ​​are practically identical; the size of the cache does not seem to change in this area. Accordingly, one can assume that the SLC cache in our 512 GB model has a size of 12 GB. This will be correspondingly smaller for smaller models. It is noticeable that the controller evidently empties the cache again during the further write process, whereby the write rate increases to the maximum value for a short moment every few seconds.

Random reading