Testing Gloway Premium 1 TB SSD on the once top-end Silicon Motion SM2262EN platform

Many buyers, faced with limited budgets when searching for a high-quality SSD or not finding suitable products in local stores, turn to shopping from AliExpress and other stores offering goods from China. Often these products are cheaper than similar ones from local stores, which becomes an attractive reason for thrifty shoppers.

However, this solution has its drawbacks. For example, guarantees provided in theory may not always be realized in practice. The main problem is that the product is bought “pig in a poke” — its quality and characteristics often remain unknown until the moment of receipt. Even when choosing branded products, there is a risk of getting a counterfeit or a discrepancy with the declared characteristics.

Some manufacturers may change product configurations without notice, which may cause discrepancies with the product description. Even long-lived companies, such as Netac, are no exception, sometimes subject to such situations. In general, the risk of buying from abroad is high, and searching for more reliable branded goods can end up being close in price to local offers or even more expensive.

Despite the risks associated with buying from abroad, some people manage to find good options at low prices. However, in general, if you try to save too much, the risks can be unacceptably high.

Realizing that the attractive SSD offer is significantly different from the standard price and promises superior performance, we, as fishing enthusiasts, could not resist the discount. This is Gloway's Premium Series SSD that promises impressive specs including a DRAM buffer, a five-year warranty, and a high TBW (transmitted write capacity) of 800TB. Unlike the “basic” series, this SSD is more interesting thanks to the use of the rare Maxio MAP1001 controller.

At that time, on the Moscow market for similar money it was possible to find SSDs with a capacity of a terabyte, but they were usually devices with low performance using a SATA interface. At best, these were SSDs with budget controllers and TLC memory, and most likely with QLC memory, which introduced certain limitations on performance.

Having agreed to this deal, as a typical buyer, our goal is to evaluate whether this SSD's promised performance is truly what it claims.

We must pay tribute to the manufacturers from Shenzhen — they have learned to package their products beautifully. At the same time, recently brands often save on packaging sizes by switching to compact blisters. And here is a normal, strict box...

...with a funny content: in addition to the SSD itself, there was not only a mounting bolt, but also a small screwdriver for tightening it.

The SSD itself is ordinary — as has been said more than once, it is generally impossible to come up with anything in the case of the M.2 2280. The only thing worthy of attention is the double-sided installation and the inscriptions on the board, suggesting the use of a Silicon Motion SM2262EN controller.

Well, we discovered that this SSD is actually an improved version of the SM2262, the company's top-end PCIe Gen3 solution that is slightly faster than the previous SM2262. This SM2262 controller is the direct predecessor to the SM2264 for the PCIe Gen4 interface, which recently pleasantly surprised us. Both controllers are powerful eight-channel solutions (for their development time, of course).

An SSD cannot do without DRAM — here it is represented by a full gigabyte and two chips, which speeds up its communication with the controller. The flash memory, labeled similarly to the original from Intel, is a fairly old 64-layer TLC with 512 Gbit die (B17A), whose best days passed about three years ago. This configuration is reminiscent of the top-end SM2262EN-based models of the past, except that in capacities of terabytes or less, the B16A with 256 Gb chips was more often used to better load all eight controller channels. In this model, the crystals are interleaved only twice, which may slightly limit performance.

It is interesting to note that in this model the MAP1001 controller is found with 64-layer YMTC memory with 256 Gbit crystals, but this modification is no longer available since spring.

Understanding that today it is customary to look for SSD speed characteristics on other interfaces is not surprising. However, the stability of the configurations is not a priority, since although they are all approximately the same class, fans want more precise tuning. For example, a temperature sensor locked to 46°C may be a disappointment not only for him.

The SSD components in all cases are relatively old, which is a minus, since it is not entirely clear where they come from at the moment. Despite this, at a low price, as already mentioned, such an SSD can become a worthy workhorse, superior to many modern budget options, including not only Chinese ones.

It is difficult to assess the durability of such a device in advance without decent foresight. This is a typical situation not only for this segment, but here this uncertainty is most clearly manifested. Therefore, without relying on foresight, it is better to just see how this SSD can work in practice.

Our test bench uses an Intel Core i9-11900K processor and an Asus ROG Maximus XIII Hero motherboard based on the Intel Z590 chipset. This allows you to connect the SSD to PCIe Gen4 “processor” lanes or to PCIe Gen3 “chipset” lanes. The first option is the most suitable for modern SSDs, allowing them to operate at full capacity. However, the «compatibility» mode is also of interest. The PCIe chipset controller in this form was introduced back in Intel 100 series chips (in 2015) and has not fundamentally changed since then. Therefore, the results in this mode will be similar for the owner of an older computer if he decides to install a modern SSD there.

This makes sense because modern interface-limited SSDs often outperform previous flagship models. That is, PCIe Gen4 is not the only advantage of the new SSDs. Sometimes they have enough other features to achieve outstanding performance. However, to understand how significant this advantage is, testing is necessary. In this case, we will focus solely on compatibility mode, since the main device does not support the higher interface, and this will allow us to compare all SSDs under the same conditions.

Understanding that the Silicon Power XD80 was recently tested on a Silicon Motion SM2262 controller (the SM2262EN's slightly slower cousin) with 96-layer Intel memory inspires a desire to evaluate what might be lost due to more limited parallelism. However, it is interesting from a research point of view. In the conditions of the second half of 2023, it is time to compare such models with more modern alternatives.

In this case, we decided to compare the initially inexpensive TeamGroup MP44L 1 TB on a four-channel bufferless Maxio MAP1602 controller with the Samsung PM9A1 of the same capacity. PM9A1 is an OEM analogue of the popular Samsung 980 Pro, but a higher class. However, the cost of the PM9A1 is at a similar level to the price of Gloway Premium, which makes it more attractive for users of systems with PCIe Gen4 support.

Owners of older computers often specifically look for SSDs for older platforms, believing that they fit together better, and prefer not to pay extra for Gen4 when it is not available. Despite this, it has been repeatedly demonstrated that the benefits of the new controllers are not only and not so much in the interface, which means that this practice is also worth testing.

Today, despite the fact that two of the four SSDs support PCIe Gen4, we will not use this feature and put all devices on an equal footing for comparison.

As you can see from the graph, the top shows the Silicon Power XD80 results and the bottom the Gloway Premium results. They are almost identical, with the exception of slight differences in certain areas due to different memory organization with different levels of parallelism: this 1 TB capacity XD80 device uses only 2x interleaving, while the 2 TB version uses 4x interleaving. As a result, the data transfer rate at the bottom of the graph decreased by half, and the operating time decreased by approximately 1.65 times. However, the general nature of the graphs is the same, which shows a standard device based on the SM2262 controller (all its modifications) without any special exclusive features.

A more detailed analysis is no longer required, since both devices demonstrate similar results and follow the typical characteristics of devices based on the SM2262 controller or its modifications.

It is interesting to note that the more modern device from Team shows almost a similar graph. However, write speeds in the SLC cache can reach more than 3 GB/s, and in direct write mode the results improve to values above 1 GB/s. But at the end of the cache clearing process, the speed drops to about 500 MB/s. If the controller were more powerful, the speed of this stage could be faster.

These observations indicate that newer SSD models have some improvements in write speed, especially in SLC cache and direct write mode, but at the end of the process the speed can be limited by the controller's capabilities.

A modern top-end drive, even within the narrow confines of PCIe Gen3, presented by Samsung PM9A1, demonstrates impressive results. Even out of cache, it is capable of writing data at over 1.5 GB/s without any apparent penalty for «flushing» the cache at the end. This confirms the statement that last year's top is, of course, a quality product, but it remains the past. Modern products may be superior, providing better alternatives even though they are more affordable.

Current low-level benchmarks, including CrystalDiskMark 8.0.1, are limited in their ability to test SSDs due to SLC caching. They provide performance information, focusing primarily on performance within the cache. However, the data provided by manufacturers about device performance is also limited to these limits. Therefore, checking actual performance is always useful.

The point of caching work is to enable devices to make the most of their cache in real-world environments. This allows you to demonstrate high data transfer rates despite decreasing memory costs. As a result, work on caching is aimed at creating devices that are able to demonstrate consistently high performance in real-world use.

Current low-level benchmarks, including CrystalDiskMark 8.0.1, are limited in their ability to test SSDs due to SLC caching. They provide performance information, focusing primarily on performance within the cache. However, the data provided by manufacturers about device performance is also limited to these limits. Therefore, checking actual performance is always useful.

The point of caching work is to enable devices to make the most of their cache in real-world environments. This allows you to demonstrate high data transfer rates despite decreasing memory costs. As a result, work on caching is aimed at creating devices that are able to demonstrate consistently high performance in real-world use.

Also significant — the SM2262EN is indeed more powerful than the original SM2262, which once mattered. And today such SSDs can sometimes compete with budget models on an equal footing. But without demonstrating any miracles, they have fallen in price for a reason.

It's even more fun with recording. The weak point of bufferless budget controllers is that it is more difficult for them to modify the address translation table. Namely, the speed of performing this operation in such scenarios can be a bottleneck. Data can be written anywhere — the main thing is to correct the addresses correctly. Models with DRAM keep a considerable part of the address table in DRAM (or even all of it fits there); bufferless controllers have to immediately (or almost immediately) write modifications to flash memory. This is slower — and limits the number of operations in «saturated» mode. What could be important — but in practice, no long queues simply have time to appear, so one cannot even hope that this difference will manifest itself in the operation of the application software in any way.

Such operations have a much greater impact on the speed of real software than the previous ones: “long” queues have nowhere to come from in practice — but blocks other than 4K bytes are very common. The number of operations per second on «large» blocks is slightly reduced, but they themselves are larger — so the resulting speed in megabytes per second is higher. Therefore, whenever possible, everyone tries to work this way. This also has a beneficial effect on bufferless controllers — after all, they often have to “climb” into the flash to translate addresses and when reading too. Fewer such operations mean less impact of this process on productivity. Therefore, the old top platforms do not have a very good time in such conditions.

And when recording, everyone is equal, just like in a bathhouse. Or almost equal. It would be most correct to say that modern tops are not far from yesterday’s (or even not at all), modern budget tops are not far from yesterday’s budget tops, but in general, in such a crush, everyone can mix up. The only obvious outsiders in such a competition may be SATA drives, which the interface itself will not let go high. But today we don’t have those.

Mixed mode is also important — after all, in reality (and not in test utilities) it rarely happens that for a long time data has to be only written or only read. Especially in a multitasking environment — and given the rich inner life of modern operating systems. But we don’t see anything new here—everything is predictable. It would seem that the presence of a DRAM buffer should have a beneficial effect — but this can only be noticed on minimal blocks. And in general, the data locality in this benchmark is too high. In practice (everyday — and not in a database server), however, too. Why manufacturers are gradually using more and more bufferless controllers — they are cheaper and do not cause visible problems. Users have a choice of what to buy while they have it. But when it's formulated as an old top with DRAM or a new low-cost controller without, there's too little incentive to go the first route.

Despite impressive results in low-level utilities, such write and read speeds are often difficult to achieve in practice. This is due to several factors. For example, utilities like CrystalDiskMark work with small pieces of data within a single file, which often results in testing information residing in the SLC cache.

However, the actual operations of writing data to an SSD are much more complex. For example, writing a single file not only writes to the cache, but also changes various system structures of the file system, such as MFT and journals. This results in the need to write data not only sequentially to one location, but also to different areas of the disk, which affects performance.

To more realistically evaluate SSD performance, tools such as the Intel NAS Performance Toolkit are often used. It allows you to test not only the cache, but also performance on devices with a small amount of free space remaining, which is closer to real operating conditions.

Despite impressive results in low-level utilities, such write and read speeds are often difficult to achieve in practice. This is due to several factors. For example, utilities like CrystalDiskMark work with small pieces of data within a single file, which often results in testing information residing in the SLC cache.

However, the actual operations of writing data to an SSD are much more complex. For example, writing a single file not only writes to the cache, but also changes various system structures of the file system, such as MFT and journals. This results in the need to write data not only sequentially to one location, but also to different areas of the disk, which affects performance.

To more realistically evaluate SSD performance, tools such as the Intel NAS Performance Toolkit are often used. It allows you to test not only the cache, but also performance on devices with a small amount of free space remaining, which is closer to real operating conditions.

Multi-threaded mode allows everyone to show what they are capable of — but within such a framework, it quickly runs into interface limitations or around them. So close that minor differences cease to have practical significance. But from a research point of view, sometimes interesting nuances are visible. For example, SM2262EN allows the Gloway SSD to overtake the Silicon Power XD80 in single-threaded mode — in the second, the slightly slower SM2262. But in multi-threaded, different interleaving played a role: both SSDs have memory crystals of the same size, but in the XD80 there are twice as many of them. The differences in both are small, but once again demonstrate to us the importance of all factors affecting speed. Unless, of course, there are some requirements for it that force you to pay attention to a few percent of performance.

As long as the size of the SLC cache is enough for everything, all controllers can show all their capabilities. But once again it is clear that the SM2262/SM2262EN could be considered outstanding only in those days when it was the developer’s top solution — and the competitors’ analogues were about the same. Now there is no wow effect — modern models, and despite the limitations of the interface (and they have an effect here, despite the absolute speed values), are still noticeably faster. If there is not enough space in the cache (it is limited in size for all participants), this looks especially harsh. But, again, what else can you expect? More than a gigabyte per second looked very good in those days and under such conditions, but now it’s not so good.

The algorithms for working “inside” the drive become the same as in the previous case, so the problems are the same. If, of course, we consider that there are such, and they are considered as such by a specific user. It is worthy of attention that the Gloway Premium SSD, when necessary to push data into the main array, “deflates” much more than the Silicon Power XD80. But the reason for this has already been indicated — the alternation is different. At the time when the platform was relevant as a top-end solution, terabyte units were no slower (if not faster) than two-byte units due to the use of memory crystals of different capacities. But, as we see, the burden of knowledge of that time is not worth carrying to this day.

The most obvious difference is between generations. Modern controllers are initially designed to quickly move large amounts of data — primarily to implement the capabilities of a faster interface, but the compatibility mode also suffers a little for the company. And yesterday’s top is, first of all, yesterday’s, and then already the top.

For hard drives, the “sequence” of access is vitally important, but for SSDs there is no particular difference between these scenarios for a long time. On the other hand, it is still more complicated than the first one — so here it is more difficult for new platforms to break away so radically from the “old guys” of a decent (for their time) level. And the latter by themselves can already cost even less than a modern budget budget (with a lot of reservations, of course, but they can). Moreover, the traditional remark that potential performance is not so easy to implement in practice (when at all possible) has also not been canceled.

Currently, one of the best comprehensive benchmarks for assessing storage performance is PCMark 10 Storage. This benchmark includes several tests that cover a wide range of use cases: from operating system loading to data copy operations, including both internal and «external» files. Among all the tests in this set, the most informative and useful is «Full System Drive». The remaining tests can be considered as its subsets, but, according to many, they are less interesting and important for a full assessment.

PCMark 10 Storage provides accurate, real-world throughput and latency measurements for real-world workloads. Averaging these metrics across different scenarios and then normalizing them to a total number can be somewhat synthetic, but this approach provides a more comprehensive assessment of SSD performance under different usage conditions. Current benchmarks are somewhat synthetic, but are closer to real-life use cases than most other tests.

Of course, it is already difficult to assess the increase in drive performance without using special tools at this level. Performance gains come from developers being able to make these changes, and there may be a point where growth slows down, as happened with hard drives that have kept pace with capacity increases and software demands. Some long-time HDD users attribute the slower performance of operating systems of the previous decade, when in fact the speed limitations of hard drives were already noticeable 20 years ago, there was simply no means of comparison with faster alternatives.

SSDs appeared on the market earlier, but due to high prices for flash memory, they did not immediately become popular. However, the reduction in prices for flash memory by orders of magnitude immediately led to a noticeable effect. Manufacturers are now able to not only lower prices and increase device capacity, keeping demand for high-capacity drives high, but also improve performance without much effort.

Old platforms were designed for outdated memory, which is more expensive to produce than new technologies. Therefore, such old platforms are no longer of interest to large manufacturers, although they may still be in demand for niche solutions if a sufficient number of components can be secured.

The decision to buy SSDs from unknown manufacturers on AliExpress is causing controversy among buyers. Some buy such products without engaging in disputes, others do not consider this option at all. Our testing illustrates one reason for this uncertainty: although we were hoping to use a different platform, what we got was not what we expected. From the average buyer's perspective, there may be no difference between the platforms, and it may even be preferable to use a certain type of platform. However, there is uncertainty regarding the predictability of such drives and the origin of legacy components. It is unclear how reliably a drive purchased from AliExpress will work in the future, and how problems can be resolved if they arise.

It is important to make a purchasing decision by weighing all possible risks and benefits, and not focusing only on price. While a particular drive may be selling for a bargain price, this may not be due solely to its quality.

Regarding the research, we looked at the old platform from a new perspective. We discovered that changes had occurred, and previously manufacturers tried to improve the performance of terabyte drives using memory with 256 Gbit crystals. However, today it is difficult to expect such an approach. High-performance solutions of this drive no longer meet modern standards, but you can try to save money with its help, although this may entail risks.