• Home
  • Blog
  • Testing SSD Kingston KC3000 2 TB: yesterday's top in the fastest configuration

Testing SSD Kingston KC3000 2 TB: yesterday's top in the fastest configuration

31.01.2024 08:07

Historically, we have tested four SSDs using the Phison E18 controller. Two of them were the highest performance version based on Micron's 176-layer B47R TLC memory, but they were all just 1TB in capacity. In the past, under the dominance of hard drives, capacity did not receive much attention because the performance of models of the same line but of different capacities tended to be similar.

However, over time, the situation has changed, especially with the advent of “trimmed” platters, which made the dependence of SSD performance on their capacity non-linear. This feature does not exist with hard drives, where even with an increase in the number of read-write heads, only one is always working at a certain point in time.

Solid state drives, having no mechanical components, avoid this problem. Instead, operations can be efficiently parallelized across memory channels and banks. However, despite this, the non-linear relationship between performance and capacity remains relevant, and maximum speed is not always achieved at maximum capacity.

In the past, 1TB models often demonstrated superior performance compared to smaller capacities, but with the advent of higher-capacity dies such as terabit, interest in higher-capacity models has increased again. Manufacturers may also find benefit in increasing capacity, even if users do not require as much capacity. This is because larger SSDs become relatively cheaper per gigabyte.

Benchmarking can be tricky because different capacities and memory technologies can impact overall performance. For example, comparing the SSD on the Phison E18 controller to other models such as the Samsung 990 Pro or WD Black SN850X may not be representative due to differences in testing conditions.

Overall, SSD technology continues to evolve, and new generations of controllers such as the Phison E26 are already being introduced with PCIe Gen5 support. Despite this, older models, such as the Phison E18, remain interesting and affordable options for users, especially when taking into account their reduced cost against the backdrop of new releases.

Kingston KC3000 2 TB

Let us emphasize a feature of the Kingston KC3000 and Fury Renegade SSD series, which was previously noted when reviewing the terabyte modification: in these devices the memory operates at a full speed of 1600 MT/s per channel, while many other manufacturers are limited to 1200 MT/s. And although there were assumptions about the potential impact of this factor on the speed of operation, in fact, 1200 MT/s is enough to fully realize the capabilities of the PCIe Gen4x4 eight-channel controller. However, the mentioned 1600 MT/s is technically considered “correct”, and for the sake of maximum correctness it is preferable to stick to the manufacturer’s specified values.

Strictly speaking, the actual manufacturer of all such SSDs is the same: they are produced under the direct control of Phison and then supplied to partners. But the labeling may give the impression otherwise. For example, Kingston has its own DRAM (actually made by SK Hynix), and their flash memory is also «proprietary» (although it is actually sliced and packaged under Kingston's control, not Micron itself). Such nuances have become commonplace in today's tech world, where companies such as Seagate and Kioxia also supply Phison controllers with their own branding. In this case, Kingston has a greater level of its own labeling.

Let us emphasize a feature of the Kingston KC3000 and Fury Renegade SSD series, which was previously noted when reviewing the terabyte modification: in these devices the memory operates at a full speed of 1600 MT/s per channel, while many other manufacturers are limited to 1200 MT/s. And although there were assumptions about the potential impact of this factor on the speed of operation, in fact, 1200 MT/s is enough to fully realize the capabilities of the PCIe Gen4x4 eight-channel controller. However, the mentioned 1600 MT/s is technically considered “correct”, and for the sake of maximum correctness it is preferable to stick to the manufacturer’s specified values.

Strictly speaking, the actual manufacturer of all such SSDs is the same: they are produced under the direct control of Phison and then supplied to partners. But the labeling may give the impression otherwise. For example, Kingston has its own DRAM (actually made by SK Hynix), and their flash memory is also «proprietary» (although it is actually sliced and packaged under Kingston's control, not Micron itself). Such nuances have become commonplace in today's tech world, where companies such as Seagate and Kioxia also supply Phison controllers with their own branding. In this case, Kingston has a greater level of its own labeling.

The differences between the KC3000 and the Fury Renegade come in several ways. In the second line, the reserve area has been increased, which leads to an increase in the “allowed” recording volume (TBW) from 400 to 500 TB for every conventional 500 GB of capacity. However, this change comes with some compensation: for example, the terabyte KC3000 has a capacity of 1024 GB, while the similar Fury Renegade is only 1000 GB. The two-terabyte models also have a slight difference: 2048 and 2000 GB, respectively.

It is also worth noting that Fury Renegade is positioned for gamers, enthusiasts and advanced users, so this series includes modifications with a radiator for more efficient cooling. While the KC3000 is equipped only with a foil heat spreader, leaving cooling issues to the discretion of the user.

Thus, when choosing between them, you should consider not only the performance parameters, but also the design features, purpose and presence of the heatsink in the Fury Renegade, which can be a deciding factor for gamers and enthusiasts.

Taking advantage of a successful sale, we purchased a terabyte SSD, which turned out to be slightly younger in release date compared to previous models (15th week of this year versus 20th week of last year). However, even in a year there have been no changes in the firmware, which makes the comparison absolutely correct. This copy is an improved version of the Kingston terabyte, and finally we can close a couple more blind spots in our review.

However, it is worth noting that there is information that not all Kingston KC3000 available on the market are identical to the model described. With the loss of the formal status of a top platform, manufacturers have an incentive to reduce the cost of SSDs based on it. In this case, this led to the use of 112-layer Kioxia BiCS5 memory. These configurations, although familiar to us since the end of 2022, are not even competitive in speed with previous versions of the platform using Micron B27B memory, not to mention the later version with Micron B47R. However, their production costs are lower, which can be beneficial for manufacturers. For buyers, this may be an unpleasant surprise, since the configuration changes, but the name remains the same.

It is unclear how widespread this trend will be. Please pay attention to this point when choosing, so as not to unexpectedly receive a different configuration. It may be worth considering alternative options, given that even less expensive models can offer better features and prices. It should be remembered that the performance of the latter is already excessive for most application software.

Testing

Samples for comparison

Of course, we need a terabyte SSD model KC3000 to evaluate the platform's capacity scalability. We've also included the three top-end SSDs mentioned above for a full comparison: Adata Legend 960 Max, Samsung 990 Pro, and WD Black SN850. The Adata Legend 960 Max uses the same memory, but a different controller (Silicon Motion SM2264) and different SLC caching algorithms, which makes the comparison even more interesting. Samsung and WD present modern high-end models, released much later, but taking into account the influence of the Phison platform, which successfully competed with their previous models. In this case, these four SSDs will provide a complete overview of the current top-end storage market.

Filling with data

The write speed to the SLC cache hardly increased, since it was already close to the maximum capabilities of the controller using this memory. What's more interesting, however, is the increase in speed outside the cache by an additional 500 MB/s, effectively increasing the overall speed by one and a half times. As a result, the total work time increased disproportionately to the amount of work performed. This happened despite a suboptimal caching strategy for writing large amounts of data, where all cells are first written in single-bit mode, and then the data has to be erased and rewritten.

Why do manufacturers choose this strategy? This is due to the fact that in practice recording volumes are often limited, especially in mass scenarios when the computer has only one drive for all tasks. A large cache ensures efficient data processing: quickly write and move on to the next stage. The faster the recording, the more time you can “sleep” until a new request arrives. This operating rhythm is standard for SSDs in personal computers.

If the write load were constant and continuous, disabling SLC caching would be preferable, as is common practice in server drives. However, in this use case, drives do not show impressive results in performance peaks — there is always a price for each solution. The second pass of recording “over garbage” perfectly demonstrates the high-speed capabilities of the combination of controller and memory. This test is scientifically important.

In particular, it shows that Micron B47R memory remains very fast when used correctly. Eight-channel controllers such as the Phison E18 or Silicon Motion SM2264 extract speeds of up to 3.5 GB/s from it. Even the new WD platform shows lower speeds, and the Samsung 990 Pro was a disappointment in this aspect. However, it should be emphasized that these results relate solely to this test and are not indicative of practical performance. Actual performance in practical scenarios depends on many factors, including the controller and SLC caching algorithms, so it is better to evaluate them under more realistic usage conditions.

Note that according to the logic of SLC caching and the achievable speeds, the Kingston KC3000 demonstrates close performance to the WD Black SN850X. However, in «standard» mode, the SN850X is ahead of the KC3000, likely due to the higher performance of the controller. This helps compensate for even the slower memory speeds found on the SN850X. Compared to these models, the Phison E18, even in new configurations, does not show its full potential. However, as already noted, the original WD Black SN850, which we had to compete with almost until the end of last year, showed similar performance, but was noticeably slower.

The Adata Legend 960 and Samsung 990 Pro use completely different SLC caching algorithms. The strength of the first is the high speed of direct writing over large areas after the cache, which makes it little sensitive to cache misses. However, it also has the downside of a higher chance of recording errors compared to Kingston and WD. The Samsung 990 Pro, as noted, is not as effective in such scenarios. However, its powerful controller pushes it forward in random addressing operations, delivering high performance in these scenarios.

Maximum speed characteristics

Low-level benchmarks in general, and CrystalDiskMark 8.0.1 in particular, have long been victims in the unequal fight against SLC caching. They can only measure cache performance, leaving out the actual performance of the device. However, the information provided by manufacturers about device performance is also limited by these limits. Thus, testing devices using such benchmarks always remains useful. Work on caching is being carried out precisely in order to use the cache as often as possible in real conditions and demonstrate high speeds, despite the reduction in memory costs.

The “full-size” modification has noticeably increased in writing and reading at the same time as writing. However, these are not improvements, but simply corrections to the situation. This is how it should be — it’s just that the terabyte capacity could not yet be deployed to its full potential.

In this scenario, on the contrary, no significant changes occur. Performance is largely limited by the controller, and the impact from memory is almost negligible. Even if there is something wrong with the memory, it can be masked by the SLC cache, which Phison controllers actively use to speed up read operations.

When performing write operations, more pronounced progress is noticeable, and this, however, is expected. Improving the internal parallelism of the memory array allows for more efficient optimization of these operations. The Phison E18 continues to excel at this task. However, unfortunately, these capabilities still remain in little demand in software.

Such operations are in great demand — but not in such quantities, roughly speaking. However, to some extent this is for the better — otherwise we would have to regret that the performance of the two-terabyte device has decreased altogether.

And when recording, it increased. In exact accordance with the theory.

Therefore, here once again you can be happy for the Samsung 990 Pro. And you can do it for everyone else — in this segment such indicators have been achieved that their improvement or deterioration (within reasonable limits) no longer matters. Because in practice you won’t be able to notice it.

Working with large files

However, despite the impressive results in low-level utilities, it is not always possible to achieve such speeds in practice. This is due to the more complex nature of real-world workloads. For example, CrystalDiskMark operates on small pieces of information, often working within a single file. Under such conditions, it is almost always guaranteed that the information being tested is located in the SLC cache for the entire testing time. In addition, there is no distraction from file system overhead operations such as MFT modification and journals that are a normal part of real-world writes, and writes occur not just to one location sequentially, but to different locations, sometimes in small blocks. For a more realistic assessment of SSD performance, it is preferable to use tools such as the Intel NAS Performance Toolkit. This tool allows you to test the device not only on an empty disk with maximum cache sizes, but also in conditions where there is almost no free space — which is closer to the real use situation.

Working in one thread is the most common, but also the most demanding scenario. However, for modern controllers this is a much less difficult task compared to their predecessors. Although the E18 is not an absolute leader in this aspect, its performance looks quite acceptable. Increasing the SSD capacity slightly improves the speed of reading data that has already been ousted from the cache — which, in general, remains relevant. However, the increase itself turns out to be small.

But in multi-threaded reading, Kingston KC3000 is the fastest. And with 2 TB of capacity, both speeds increased. One problem is that you rarely have to read at a high pace in several threads sequentially.

With the increase in capacity, the speed also increased, allowing it to take the first line of the rating, regardless of the current state — the test file almost always ends up in the cache. However, unlike read operations, none of the devices manage to completely saturate PCIe Gen4. However, we have already discussed the reasons for this phenomenon — the functions of working with files in operating systems, including Windows, have remained largely unchanged since the days when gigabytes per second were a distant future. And this is not so long ago from a historical point of view — just a decade ago. Thus, to fully unlock the potential of modern SSDs, it is necessary not only to revise the application software, but also the system software.

Regardless of how requests are generated, the system driver first places them in one queue, after which the SSD carries out parallelization within itself. Why there is a difference between these scenarios and what it is can be explained by the fact that «external parallelism» provides additional opportunities for optimization.

Despite its advanced age, the Phison E18 remains an excellent choice for working with large amounts of data, provided that quality memory stands in its way. In typical «canonical» configurations, where everything is configured optimally, performance is impressive. However, it is worth considering that versions using BiCS5 do not achieve such outstanding results. At the same time, there are practically no scenarios where the performance of this class of SSD becomes a bottleneck, since often working with data involves not only reading and writing, but also processing it, which depends on the performance of other computer components.

Previously, an important criterion was the difference between sequential and random access of data in mechanical drives. Manufacturers focused on sequential access because it was much more efficient. However, in practice, random access turned out to be more relevant. This was one of the reasons for abandoning mechanical drives in favor of SSDs in cases where performance is important, since with SSDs the difference between these scenarios is not so great. Although there are slight differences between the top models, the Phison E18 is still among the best platforms. In some scenarios it may lose some of its credibility, but in others it remains on top.

Comprehensive performance

Currently, PCMark 10 Storage is the best comprehensive benchmark for assessing storage performance. In our previous review, we highlighted that not all three tests included in this set are equally useful. The most informative is the “full” Full System Drive test, which includes almost all the main usage scenarios: from loading the operating system to copying data (both internal and “external”). The other two tests are only subsets of it and, in our opinion, are less “interesting”. This test is useful not only for accurately measuring throughput when solving practical problems, but also for estimating the resulting delays. Averaging these metrics across scenarios and then reducing them to a single number can be a bit synthetic, but given the lack of more realistic «overall» estimates at the moment, this leaves this testing quite useful for assessing drive performance.

For almost three years now, the SSD market has been undergoing significant changes, especially in the high-end segment, where the pace must be constantly increased to even stay in place, and progress requires running even faster. Although the Phison E18 platform remains a high-performance platform, competitors have already released updated models that perform more efficiently. This is especially noticeable in the mode of lack of free space, when SLC caching cannot fully develop. A difference is emerging and newer models are showing better performance.

However, it is worth considering this aspect from two sides. Firstly, this approach smoothes out the differences between SSDs of different capacities. Even if they have different memory speeds, it is difficult to notice in real-life use. Secondly, the concept of “slow” is very relative. Even the best SATA models achieve approximately 1000 points in this test, which is considered sufficient for many users. The best SSDs with PCIe Gen3 also fully satisfy the needs of most, falling within 2000 points. Despite the criticism expressed in the test, this level of performance remains very respectable and fully satisfies the needs of most users.

Total

Regardless of the specific product, it is clear that the approach to SLC caching in the SSD industry has changed over time. In the past, most vendors tended to tune caches so that they were not heavily dependent on regular hits and misses. Examples include the static caching schemes used in older Intel and WD SSDs, as well as in some Samsung 970 Evo/970 Evo Plus models, where dynamic SLC caching was limited and amounted to a few percent of the total capacity of the device. However, with the advent of PCIe Gen4, this approach has changed — large caches have become more common, and cache clearing algorithms have become more aggressive.

But, as it turned out, this approach is not always a panacea, and developers began to balance SLC caching algorithms. New tests show that the maximum speeds are achieved only when working with the cache, but the optimization algorithms for misses have become more efficient. An example here is the Samsung 990 Pro, where even with reduced memory speed, the controller allows you to achieve outstanding results, carrying out optimizations even with a parallel load.

However, when it comes to older models, such as the Phison E18, which was released earlier, you can notice a difference in approach. The Phison E18 is not aggressive when cleaning the SLC cache, which can have an impact on long-term performance. Controllers from other manufacturers usually clear at least the static part of the cache as soon as possible, while the E18 does not. This means that if space is limited, the E18 may not maintain high levels of performance during actual use.

On the other hand, Phison has released the new E26 and E25 controllers, which are already delivering impressive performance even with limited space and interface limitations to PCIe Gen4. These models may be expensive at first due to the premium for PCIe Gen5 support, but over time they will replace older models on the market. Interestingly, Phison has released two controllers to combat PCIe Gen5, which will likely lead to more affordable options on the market.

For buyers, these models have two advantages. First, aggressive SLC caching reduces the dependence of performance on total capacity. Secondly, the decline in popularity of the E18 platform is affecting prices, and buyers can take advantage of price cuts on previous top models. However, as always, you should be careful when purchasing, as the use of slower memory may become more common.