• Home
  • Blog
  • Testing SSD Digma Pro Top P6 1 TB on the first and so far only controller with PCIe Gen5 support

Testing SSD Digma Pro Top P6 1 TB on the first and so far only controller with PCIe Gen5 support

20.12.2023 10:17

Computers with PCIe Gen5 slots have been selling in large numbers for two years, but there are no suitable devices to fit into these slots yet. Although there is plenty of hardware compatible with different versions of PCIe, there is practically no equipment specifically designed for Gen5. This is due to several factors. Firstly, a new interface usually appears before the corresponding devices are developed for it. Secondly, computer updates have slowed down, and many systems still don't even have PCIe Gen4 support, let alone Gen5.

Most users understand that increasing interface bandwidth does not always lead to improved performance. In the early stages, many devices associated with the new interface may simply use it to indicate that they support the latest technology. This may increase the cost of the product without significant performance gains. Until new features become more relevant, users are in no hurry to upgrade computers due to new internal interfaces.

Manufacturers continue to focus on new interfaces, but at the moment neither AMD nor Intel are releasing GPUs with PCIe Gen5 support. For example, Nvidia has cut the interface to eight PCIe lanes even in more expensive chips. This is due to the fact that in older systems with PCIe Gen3, even modern video cards work relatively normally.

Intel simplifies the situation in this regard, since support for PCIe Gen5 within the LGA1700 is implemented exclusively by the processor controller. However, AMD, when introducing the AM5 to the market, emphasized that it has more lanes and more options for connecting SSDs. But in a global sense, few people are yet paying attention to new interfaces. However, there is interest in PCIe Gen5 and developers are keeping an eye on its development despite its slow adoption.

Computers equipped with PCIe Gen5 slots have been selling in droves for two years, but there are currently virtually no devices specifically designed for these slots. Despite the availability of equipment compatible with different versions of PCIe, there are practically no devices adapted for Gen5. This is due to several factors. First, a new interface often appears before the corresponding devices are developed. Second, computer hardware upgrades have slowed, and many systems still don't even support PCIe Gen4, let alone Gen5.

Most users understand that increasing interface bandwidth does not always lead to improved performance. In the early stages, many devices associated with the new interface may simply use it to indicate that they support the latest technology. This may increase the cost of the product without significant performance gains. Until new features become more relevant, users are in no hurry to upgrade computers due to new internal interfaces.

Manufacturers continue to focus on new interfaces, but at the moment neither AMD nor Intel are releasing GPUs with PCIe Gen5 support. For example, Nvidia has cut the interface to eight PCIe lanes even in more expensive chips. This is due to the fact that in older systems with PCIe Gen3, even modern video cards work relatively normally.

Intel simplifies the situation in this regard, since support for PCIe Gen5 within the LGA1700 is implemented exclusively by the processor controller. However, AMD, when introducing the AM5 to the market, emphasized that it has more lanes and more options for connecting SSDs. However, at the moment, few people pay attention to new interfaces in a global sense. However, interest in PCIe Gen5 remains, and developers continue to monitor its development.

Phison E26 as a practical implementation of the concept of the first pancake

Three years ago, Phison E16 attracted a lot of questions and criticism, especially in the context of an unconvincing demonstration of the benefits of the new interface. As the market flooded with alternative products, it became clear that not all controller problems were interface related. It turned out that the speed of execution of certain scripts does not depend on the interface, even if these are key aspects. However, Phison also had its shortcomings, such as low sequential read and write speeds, which made their E16-based SSDs less competitive. Insufficient support for fast memory interfaces also caused limitations on maximum speeds. By that time, there was not yet a wide selection of memory with high interfaces on the market. Even the next generation of Phison E18 controllers initially suffered from a lack of support for high-speed memory, and this problem was only resolved in mid-2021, a couple of years after the release of the E16. The initial stage showed that some problems remained unresolved due to limitations in the memory available at that time.

Top controllers for consumer SSDs are mostly eight-channel. Fewer channels reduce the overall data transfer rate between the controller and flash memory, which is unacceptable for high-performance devices. It is impossible to place more than eight channels on standard M.2 cards (22 mm wide). All this creates problems when achieving high speeds, especially for server SSDs, where the increased speed of the interface between the controller and the system is relevant.

Therefore, a fast memory interface becomes key. At the time of development of the Phison E16, memory with a speed of 800 megatransactions per second (MT/s) per channel was widely available. With eight channels, the controller could only handle 5 GB/s (5000 MB/s). This is significantly more PCIe throughput than Gen3x4, but less than Gen4x4, which limited the effect of moving to Gen4. When 1200 and 1600 MT/s memory became available, PCIe Gen4x4 throughput increased. But even now, available memory speeds (eg 1600 MT/s) only provide a maximum of 10 GB/s.

The throughput of each Gen5 line is twice that of Gen4 and four times that of Gen3. To use the standard 14 GB/s, you need memory with a speed of 2400 MT/s, which is still rare. Even available 1600 MT/s memory provides only 10 GB/s. Based on the Phison E26 controller, several SSDs have been introduced, with some offering slightly higher read speeds, but most are limited to 10 GB/s. Products with high read and write speeds remain problematic due to available memory limitations.

Phison E26 controllers are manufactured at 12 nm standards, which leads to increased heat generation. Under load, they can consume more than 10 watts of power, causing thermal concerns. Setting up throttling was a challenge, but thanks to optimization it was possible to reduce thermal conditions. Now SSDs work stably, but with reduced performance. Manufacturers clearly indicate the need for active cooling for controllers.

In addition, many manufacturers have released special coolers for these SSDs. Some bundle them with drives, others offer them as a separate accessory. Although a cooler is not required, this issue still needs to be resolved. Standard versions usually have small high-speed fans that cannot operate quietly even when new. Massive passive heatsinks with heat pipes have also appeared that can cope with the high thermal load of the Phison E26, but they can be too massive and bulky.

In fact, you can often come up with a quieter and more efficient solution yourself by using simple metal plates and installing the right size fan, such as 80mm. However, this will require some time and effort.

All this creates a somewhat paradoxical situation. New PCIe Gen5-enabled SSDs are already available for purchase, but it's hard to avoid the thought that this isn't quite the «proper» Gen5: it's known that speeds can be faster than current products. And, of course, they will. In addition, prices will likely go down as there will be no exclusivity.

It is also clear that not all speed characteristics will improve significantly with the transition to the new interface. New controllers and memory should certainly improve performance, but that's about it. We've seen this before with the transition from PCIe Gen3 to Gen4, and that happened less than five years ago, so the memory is even fresher.

In addition, the new SSDs are unlikely to interest users of the Intel LGA1700 platform, because the stock M.2 slots on these motherboards are limited to PCIe Gen4. Yes, top boards have two large PCIe slots, but using one of them automatically switches the video card mode from x16 to x8. And this will not attract owners of RTX 4060 Ti or lower graphics cards who are already running on x8 — it is unlikely that they will want to overpay for an SSD.

Most older AMD and Intel platforms do not support this technology either. Only the AMD AM5 platform remains, where there are corresponding slots, but so far the number of its users is small. In addition, there are not many offers of corresponding SSDs on the market either.

Digma Pro Top P6 1 TB in two incarnations

If the issues listed above are not a concern, there are already plenty of options on the market to choose from. However, it's important to remember that all of these SSDs are currently manufactured directly under Phison's control and shipped to all partners, so all of the newest models are essentially the same.

Thus, the determining factors will be not so much technical characteristics as, for example, attitude towards a particular brand or warranty issues. What you end up with is an SSD that's made in the same factory where the 2280 board features the same eight-channel Phison E26 controller, 1, 2, or 4 TB of Micron B58R flash, and 2 GB of DRAM per terabyte containers. The doubled amount of DRAM has more of an image value and was often used in Phison platforms at the start of production.

Characteristics and prices influence the choice between versions of the platform — faster or slower. The Digma Pro line includes 1 and 2 TB models, but, unfortunately, there is no 4 TB option, which is typical for most manufacturers. And although it would be better to test the 2 TB model to fully evaluate the potential capabilities of the platform, cost becomes a determining factor.

Now both modifications of Top P6, even 1-terabyte ones, are more expensive than Digma Pro Top P8 by two terabytes. Moreover, the latter is a top-end solution based on the Phison E18 controller and 176-layer Micron B47R TLC memory (by default for all Digma Pro SSDs). At the same time, the price of the “fashionable” 2 TB is almost equal to the 4 TB Top P8. From the point of view of reasonable economy, there is no rational reason to rush for progress now. A «try» approach may make sense when considering a lower capacity version.

There are currently two versions of the 232-layer Micron B58R memory, operating at speeds of 1600 MT/s and 2400 MT/s. In the case of Digma Pro Top P6, the version of a specific model is indicated on the packaging and by the full model number (last digit 6 or 4). However, it is worth noting that you should not expect double performance — even the fastest memory is used in 2000 MT/s mode. This applies not only to Digma Pro, but also to all similar SSDs on the market. Currently there is no «full speed» mode yet, and such SSDs are not yet available for purchase. If you are planning to purchase an SSD, then it is worth considering the latest version as the price difference is approximately 10%. If you just need to test, it makes sense to compare both versions to evaluate their performance with limited capacity.

The kit includes a special cooler, which, as noted above, is available as a separate product. Our opinion is that cooling in this case is necessary and mandatory. The manufacturer also insists on this requirement. Regarding the proposed cooling system, it is similar to similar models on the market. Among the advantages, we can highlight the fact that it works efficiently and is guaranteed to cope with cooling even under conditions of prolonged and intense loads on the SSD. The main disadvantage is the presence of another fan in the computer, which is also small. The new fan is currently fairly quiet, but this may change over time and the speed may increase. It is recommended that you resolve this issue yourself if it requires your attention at all.

Testing

Samples for comparison

In addition, when comparing with two top modern devices — Samsung 990 Pro and WD Black SN850, participating in testing with two-terabyte modifications, it should be taken into account that the price difference between them is quite comparable. For example, the 990 Pro costs about the same as the terabyte Digma Pro Top P6. This reflects the real cost of adopting advanced technologies, which can be quite significant. Great, now let's see what practical benefits this advanced technology provides over the already established leaders for PCIe Gen4.

It is clear that we will not use these models for comparison with Samsung and WD, since such tests have already been carried out. They are of interest to us as a tool for more accurate analysis of the new Phison platform. Regarding the choice of Seagate FireCuda 530 and Adata Legend 960 terabyte, there are some reasons. For example, in the case of Seagate, the SLC caching algorithms are quite comparable to SSDs on the Phison E26 platform, in contrast to the KC3000 from Kingston, where these algorithms differ significantly from E26 drives. As for the Legend 960, this is an example of an SSD with fast B47R memory, but with a Silicon Motion SM2264 controller. This will allow the new Phison platform to be more accurately assessed in the context of existing alternatives. By the way, there is also an analogue of the Top P6 with 1600 MT/s memory in the Adata range, called Legend 970. Such a comparison will also be interesting for fans of this manufacturer’s products.

Filling with data

The full set of graphs is too large, so it’s easy to get confused in it. Therefore, we will limit ourselves to the most important to understand. For starters, there are four SSDs on two Phison controllers and one Silicon Motion in a PCIe Gen4 slot.

The new SSDs based on the Phison E26 platform showed only a slight increase in the write speed to the SLC cache compared to the previous generation, reaching approximately gigabytes per second. However, there is virtually no difference between «fast» and «slow» modifications in this mode. It is interesting to note that the operating speed of both modifications is slightly inferior to the previous top generation. When compared with an Adata SSD based on a Silicon Motion controller, no increase in cache write speed was noticed. Overall, the performance of these SSDs was very similar, which may indicate a slight improvement in the performance of the new Phison E26 platform compared to previous models or alternative competitors.

If we consider the results of the second pass through the garbage, it becomes more clear why such dynamics occur. Firstly, it is noticeable that the Phison E26 does not clear the static part of the SLC cache without external influence, which is inherited from the E18. Secondly, the native recording speed for the B58R is lower than that demonstrated by the B47R. However, this is true for equal capacity. Considering the difference in the size of the crystals, this is logical — older terabytes worked with at least double interleaving, but now this does not happen. However, the drop in speed is not so significant due to the increase in internal parallelism. However, it still occurs and can only be compensated by increasing the capacity. If you need several terabytes, that's great. However, if a terabyte is enough, then you should carefully weigh your needs and capabilities.

In general, the compatibility mode for these newest SSDs is not at all interesting — especially when it comes to recording. And using the native interface is interesting — the opportunity to break out of the limitations of PCIe Gen4, at least when writing to the cache. And here there is at least a difference between the modifications. And not so small, although the main advantage will have to be the speed when reading data.

The second run did not bring anything new, and this was expected, since its speed depends on the memory, which remains unchanged. Moreover, the memory interface does not have a significant impact here, and even PCIe Gen2 could in theory be sufficient.

For supporters of technological progress, the conclusions are disappointing: the bottlenecks are usually not in the interface. Rather, we can immediately note where the interface is critical: when reading large amounts of data sequentially. However, the memory writing speed is still not that high. And operations with random addressing will greatly depend on the controller and only then on other parameters. Currently, it is not a situation where slow memory slows down operations much. Therefore, we end with this test scenario.

Comparing the results with Samsung and WD does not make much sense — the capacities differ there, and most importantly, the SLC caching settings are significantly different. However, these four were tested under approximately the same conditions. So we conclude (and apologize for the pun): write speed is not the main reason why you should rush to buy Gen5 drives now.

Maximum speed characteristics

Low-level tests in general, and especially CrystalDiskMark 8.0.1, have long been unable to accurately reflect actual SSD performance due to SLC caching. They mainly evaluate cache performance only. However, information provided by manufacturers about device performance is also often limited by these limits, so checking their actual performance is always useful.

Work on optimizing caching is carried out in order to use the cache as often as possible in real-life use cases and demonstrate high speeds, even with lower memory costs. In the SSD area we are exploring today, high production values are expected. These may be the highest, and it is important to understand whether these results can be replicated under more realistic use conditions.

The capabilities promised by the manufacturers were not seen in actual tests. However, this primarily raises questions for Phison. We will reserve our assessments for this SSD, at least until a larger capacity model appears — since the presented model is clearly not enough for terabyte drives.

In any case, it is clear that PCIe Gen5 is functional, and there are differences between different versions of memory. The controller has also been seriously redesigned to handle large amounts of data at high speed. This is especially noticeable in mixed use modes, where the new interface is not as significant as the controller's ability to operate at high speeds on two data streams. It's important to note that this is possible even in Gen4 mode since PCIe is a full-duplex interface.

When using CrystalDiskMark's default settings, sequential write speeds may be slightly slower, while read speeds may increase. However, the write speed will still remain higher than that specified in the SSD data specifications, and the read speed may be slightly inferior to these indicators, but only slightly, as mentioned above. The main thing is that both speeds exceed the capabilities of PCIe Gen4. It is also worth noting that the speeds of these indicators will differ between two SSD modifications with different memory.

The improvements made to the controller are noticeable here too. The move from Phison E18 to E26 is truly a significant step forward. However, competing companies are also moving in this direction, so there is no clear leadership. The overall result still remains very high. It is important to note that the interface, in this case, matters less since the bandwidth was already sufficient. Despite the slight reduction in latency, it is the characteristics of the controller that have a decisive influence. And the Phison E26 is certainly a good step forward in this regard.

When using PCIe Gen5, support increases, but record-breaking results are not initially expected due to the limitations of low-capacity configurations (and in this class, a terabyte is already incredibly small) and insufficient parallelism. Both Digma Pro SSDs begin to reach their limits as the queue depth increases, approximately around 400 thousand operations per second. However, this value is only theoretical, since in practice in ordinary computers queues do not have time to form. When using hard drives, such situations may arise, but any SSD, even not a top-end one, processes requests faster than they arrive.

Operations with different data blocks have a much greater impact on the speed of real software than previous ones. Long queues of operations do not occur in practice, but data blocks other than 4K bytes in size do occur frequently. The operations per second speed of large blocks decreases slightly, but the size of the blocks themselves increases, resulting in a higher resulting speed in megabytes per second. Therefore, most applications prefer to work with data blocks of this type. However, here we are faced with a similar problem as in the previous scenario: the controller can handle more data than a terabyte memory modification can provide. This problem affects both Adata and Kingston, although to a lesser extent. However, for the latest platform, a volume of one terabyte is extremely insufficient. This is the absolute minimum you can get.

Writing data is a simpler task. Increased bandwidth and reduced latency in PCIe Gen5 allow speeds of up to 7 GB/s on large blocks of data — which is basically unattainable with Gen4. Unfortunately, in practice such operations are used less than random read operations, but they are still important. And for testing purposes, they matter: We discovered another task for which switching to a new interface provides benefits that differ from normal sequential operations.

Mixed mode operation is also important because in real life there are rarely situations where data is only written or only read for an extended period of time. This is especially true in a multitasking environment and taking into account the active operation of modern operating systems. The problem is that all factors come into play: a powerful controller pushes devices forward, a fast interface helps it in this, but slow memory sometimes becomes a barrier. The best memory available on the market for a given platform does not always meet the requirements. In order for everything to work optimally, modifications with larger capacities, improved characteristics and the absence of frequency restrictions are necessary.

Working with large files

Tools like CrystalDiskMark don't always reflect real-life speeds. This is because CrystalDiskMark uses small chunks of information within a single file, accessing primarily the SLC cache, but also handling file system overhead operations such as MFT modification and journaling (in the case of journaled file systems such as NTFS). In practice, the writing and reading processes do not occur sequentially in one place, but in different areas of the storage. In real life, operating systems do complex work, and as a result, some of today's top-end SSDs may run into software limitations.

One tool that allows you to get more accurate data is the Intel NAS Performance Toolkit. It tests devices not only under conditions of using a cache on an empty disk, but also under more realistic conditions when the device has little free space.

We often encounter a problem when modern SSDs already provide such high speeds that the bottleneck becomes not the hardware, but the software. Operating systems are constantly updated, but basic elements of code written many years ago still exist and affect overall performance. This means that I/O performance is often limited by the speed of a single processor core. Initiatives such as DirectStorage are trying to offload the data stream by splitting it into separate processes and improving data I/O mechanisms, but this is still being developed. Updating operating systems can make a difference, but outdated code can still slow down the process.

Working in one thread is the most common (146% of cases), but also the most difficult scenario. But the problem described above does not arise here — we are still testing SSDs, and not their environment. And new market products have nothing to be proud of here — because of memory. More precisely, its lack in such configurations. If there is more, there will be higher internal parallelism. And speed, accordingly.

In the same scenario, the whole range of problems begins to play against the drives, so we don’t get any “low-level” 10-12 GB/s. But there is an effect from the transition to PCIe Gen5. Even such “non-optimal” (in terms of achieving maximum performance) configurations exceed the limitations of Gen4. More recently, the main problem manufacturers solved was to achieve them.

It would seem that PCIe Gen4 is still far from reaching the ceiling here. However, in this mode, both SSDs on the Phison E26 are only at a level slightly higher than the top drives of the “previous” wave, and when moving to Gen5 they add another 300 MB/s. Why so few? And it’s difficult to realize potential speed capabilities when the operating system itself and its functions become a limiter. To move on, you need to deal with it first.

Immediately there is a feeling that the software is “confused” — since the speed has not increased, but exactly the opposite. Moreover, the version with fast memory turned out to be slower than with slow memory. So it has been experimentally established that in the range of 4-5 GB/s miracles begin to happen. Can modern SSDs theoretically write data faster? These are their problems.

If we consider the results of new products in isolation from all others, then everything is more or less logical. If without it, then the speed is too low. On the other hand, as already said, now everything is tied up in such a dense ball of software and hardware that sometimes it is not possible to find a specific “plug”. Several gigabytes per second are achievable for modern SSDs, but they may simply be prevented from achieving anything more.

It is possible that external factors may also influence this scenario. Because of this, the high level of performance found in high-end SSDs is probably not available in budget products. However, this allows you to differentiate these products and set them apart from more expensive models that become so fast that they exceed expectations that programmers had not previously thought about.

Comprehensive performance

Currently, the most useful and comprehensive tool for evaluating storage performance is PCMark 10 Storage. We discussed this tool in our review, noting that of the three tests included in the suite, the most informative is the «full» Full System Drive test. It covers a wide range of use cases, from booting the operating system to copying data (internal and «external»). The other two tests represent only part of this comprehensive test and, in our opinion, are not as interesting.

PCMark 10 Storage allows you to accurately measure the performance of a drive, not only in terms of speed, but also the latencies that occur when performing various tasks. While averaging these metrics may be a bit synthetic, it provides a more general estimate that is closer to real-world usage. At the moment, this is one of the most realistic ways to evaluate the performance of drives in various use scenarios.

An event has appeared: finally someone has surpassed the Intel Optane SSD 900P, released in 2017. However, this came with restrictions. The Optane SSD operated successfully without caching, which cannot be said about modern TLC drives, where caching plays a key role. As long as you have access to the cache, everything is fine, but problems begin when you have to operate on main memory, especially if it is not as fast as required. Models with higher capacity are likely to provide more consistent results.

However, the main problem is not only this. Modern drives can no longer solve problems that were previously relevant. For example, for hard drives, the main difficulty was random addressing operations. Even the best of them had a limit of about 1000 operations per second, but in an «office-browser» type scenario, sometimes you need to provide 2-3 thousand operations per second. Hard drives simply can't handle it, which results in slower performance. Any even the simplest SSD solves this problem. However, the most affordable models may have new problems that are unexpected for inexperienced users, for example, low write speeds, including sequential ones. While this is often hidden by SLC caching, if it fails it will also cause slowdowns when copying large files. However, already in the middle price segment this problem disappears, which means that from the point of view of the average user these problems no longer remain. They still exist, of course, but they are fewer in number. Sometimes just numbers in benchmarks can cause positive emotions, but this is not as significant as before.

Total

Why does SSD performance continue to improve even though many people no longer know how to use it effectively (which becomes a problem not only for users, but also for software developers)? The main reason is opportunity. The main trend is to reduce costs. In the budget segment, the task most often is not to increase speed, but at least to maintain the current level of speed when releasing a new generation. And this task is usually successfully solved, especially at higher levels. Therefore, every year users receive more capacious and faster SSDs for the same money (and prices in the budget segment continue to fall). This requires mastering faster interfaces, which sometimes become a bottleneck in themselves. Manufacturers of top-end devices have to aim higher, as they are being squeezed by mid-range competitors. This is the only way to survive in this market.

However, all this success comes at a price. Exclusive benefits, even if they are nominal, are expensive and are sold at a noticeable markup. The cost of such devices is much higher than that of mass models. Some problems cannot be solved simply by money. For example, the Phison E26 in conjunction with the Micron B58R can load full PCIe Gen5x4 when reading, but the memory suitable for this is still only enough for a limited number of devices, and not for mass production. Sales have to be made on slower modifications, which does not contribute to their popularity. Especially considering the high prices and cooling problems. This also limits the scope of application of such devices. Initially, they are actually intended only for AMD AM5. Unsurprisingly, the E25 was released at the same time as the feature-limited E25, which was used in the recently released Crucial T500. But most likely the E25 is no different from the E26, with the exception of blocked PCIe Gen5 support.

If we ignore global problems and philosophical questions of necessity and sufficiency, then everything is simple. SSDs with PCIe Gen5 support exist and you can just go and buy them at any store. If you have such a desire and financial capabilities, but this applies to any equipment. And they are available from many manufacturers, although in fact there is only one “manufacturer” here — Phison itself. But not everyone has both versions (“fast” and “slow”) — so Digma Pro Top P6 stands out compared to some. In general, this is the most modern solution on the market. Even if Phison’s victory has to be counted only on points — due to the absence of at least one of the opponents. And let all such SSDs have common not only achievements, but also problems. At least there is something to buy. Whether it is necessary to purchase is a separate question. At a minimum, you can see numbers in benchmarks that are unattainable when using PCIe Gen4.