Kioxia Exceria Pro SE10 M.2 NVMe SSD (PCIe Gen 4.0)

NVMe

NVMe is also known as Non-Volatile Memory Express or the Non-Volatile Memory Host Controller Interface Specification. The idea behind NVMe is to improve the storage stack by optimizing the way an application accesses a Flash device. NVMe cuts corners by removing components within the I/O path like that good old-fashioned RAID controller. NVMe leverages PCIe as transport media which offers high bandwidth and a direct path to the host’s CPU and memory. This, in turn, removes another potential bottleneck, the limited bandwidth of SAS or the latest SATA3 connection. NVMe makes use of your PCI-Express lanes, and as you know… the latest gen 3 PCIe lanes are fast, and thus lift the bottleneck that SATA3 has. Moving from PCIe Gen 2 to Gen 3 doubles the bandwidth available to the add-on cards installed, from 500 MB/s per lane to 1 GB/s per lane.  PCIe 4.0 once again doubles that up, Obviously, the hardware you use needs to be Gen 4.0 compatible. 

So that the NVMe M.2 units use whatever they are assigned; most motherboards and units now can connect on an x4 link, so that Gen 4.0 x4 link of yours can handle a total bandwidth of 16 GB/s (bi-directional) in PCIe Gen 4.0 mode, whereas the old Gen 1.0 x16 slot would be capable of 2GB/sec (bi-directional). That bandwidth combined with fast NAND and a top-notch controller can allow for massive storage speeds. To be able to make use of the new NVMe based devices inside your PC you have to have a proper driver installed, this can be downloaded at the Samsung website. This enables NVMe devices to function within the existing operating system I/O stack protocol.

SATA 3 (6Gbps)

SATA 6G (SATA 3), the latest revision of your SATA storage unit connectors, will increase the bandwidth on the SATA controller from 3 Gbit/sec to 6 Gbit/s. For a regular HDD that is not really very important. But with the tremendous rise of fast SSD drives this really is a large plus. Typically we get 3000 Mbit/s : 8 = 375 MB/s bandwidth minus overhead, tolerances, error-correction and random occurrences. SATA 3 doubles it up, as such we get 6000 Mbit/s : 8 = 750 MB/s (again deduct overhead, tolerances, error-correction and random occurrences) of available bandwidth for your storage devices. As you can understand, with SSDs getting faster and faster that’s just a much warmed and welcomed increase of bandwidth. Put Sata3 in RAID and you’ll have even more wicked performance at hand. Most motherboards offer only two ports per controller though, so you are (for now) limited to RAID 0 and RAID 1 (stripe or mirror). 

What is NAND?

Flash memory is an electronic non-volatile computer storage medium that can be electrically erased and reprogrammed. Introduced by Toshiba in 1984, flash memory was developed from EEPROM (electrically erasable programmable read-only memory). There are two main types of flash memory, which are named after the NAND and NOR logic gates. The NAND type is primarily used in main memory, memory cards, USB flash drives, solid-state drives (those produced in 2009 or later), and similar products, for general storage and transfer of data.

NAND Types

At the beginning, memory cells stored just a single bit of information. However, the charge on the floating gate can be controlled with some level of precision, allowing the storage of more information than just 0 and 1. So you can store more with the same cell? That’s easy cost reduction. Based on such an assumption the (Multi-Level Cell) memory came into play. To distinguish them, the old memory type was called SLC – Single Level Cell. In the year 2019 we now differentiate four primary NAND storage wite methods:

1. SLC (1 bit per cell written) – fastest, highest cost

2. MLC (2 bits per cell written)

3. TLC (3 bits per cell written)

4. QLC (4 bits per cell written) – slowest, least cost

The decision of choosing between SLC, TLC, QLC and MLC is driven by many factors such as memory performance, number of target erase/program cycles and level of data reliability. The MLC memory endurance is significantly lower (around 10,000 erase/program cycles) compared to SLC endurance (around 100,000 erase/program cycles).

Toggle-mode MLC – Toggle-mode MLC is asynchronous NAND that is supposed to provide similar performance as synchronous NAND, but at a lower price. Independent testing has not verified these claims yet. Toggle-mode MLC is also known as double-data-rate asynchronous NAND. Synchronous and asynchronous NAND, based on spec sheets, look remarkably similar in performance. However, they aren’t. Synchronous NAND is more expensive than asynchronous. Sync NAND is used when performance is everything, such as with gaming systems.

TLC flash memory (triple-level cell flash) is a type of solid-state NAND flash memory that stores three bits of data per cell of flash media. TLC flash is less expensive than single-level cell (SLC) and multi-level cell (MLC) solid-state flash memory, which makes it appealing for consumer devices that use solid-state storage. The drawbacks to using TLC flash are performance, reliability, and longevity. TLC flash has lower write endurance than both SLC and MLC flash. Generally, the more bits of data the cell has, the fewer write cycles it will support. SLC memory cells can withstand up to 100,000 write cycles before failing. A 2-bit MLC memory cell can typically withstand up to 10,000 write cycles before failing. A TLC memory cell can sustain about 1,000 write cycles before failing, which is why thus far it has been limited to consumer-grade applications.

Ever since 2019 we now also have QLC NAND writes 4 bits per cell. Adding more bits per cell also has an effect on the life-span of the NAND cell, and thus that brings down the number of times it can be written. Much like TLC (Triple-level cell), many new technologies like error-correction mechanisms and wearing have increased the life-span of the respective SSDs. For example, a 500 GB TLC based SSD can quite easily manage a 300TB written before NAND cells start to die off. TLC has roughly a 1000 PE cycles, and that is the claim for QLC as well, a 1000 PE cycles. On 64-layer 4bits/cell NAND technology, companies are achieving 33 percent higher array density compared to TLC, which enables them to produce the commercially available 1 terabit die in the history of semiconductors already.

Product Showcase

The tested unit then; you should easily be able to place the M.2 unit into a compatible NVMe protocol motherboard. Most motherboard chipsets support it. You should, however, check out with the motherboard manufacturer if you have an x4 lane PCIe Gen 3.0 version with NVMe protocol support. Of course, these SSDs are backward compatible thus PCIe Gen 2.0 will work as well, however, the interconnect is halved in bandwidth per generation and that has an extensive effect on performance. The latest Windows 10 iteration has an up-to-date NVMe 1.4 protocol driver natively, so you do not need to install a 3rd party driver.  The 2TB model has 2GB of DRAM (Hynix) as cache, and 2x four NAND chips (96L) (each side). We’re quite confident that these SSDs are bought from an OEM somewhere as we recognize the sticker on that NAND chip as well as the design. The compact M.2 2280 form factor ensures compatibility with the next-generation desktop and mobile platforms that support the M.2 PCIe slot and interface. The 80 on 2280 is short for 80mm, aka, that is the length of the card and 2280, you guessed it now .. 22mm for its width. The heatsink does raise a new concern, it’s higher and very low PCIe devices could be blocked albeit there was enough clearance for graphics cards that we quickly inserted to check that out.

The Kioxia Exceria Pro 2 TB SSD is supplied in a cardboard exterior package. There’s no heatspreader or heatsink included.. 

Inside the case, you’ll spot the NVMe M.2 SSD safely seated into its protection, along with the install Guide. You can see SMT traces for even bigger versions on the backside, other than that it shows no components.

The 800TBW for this 2 TB model seems a little on the shy side. You’ll receive five years warranty; that’s should be more than plenty.  On top you can read out Toshiba TC58NC1210GSE-00-BB, though Kioxia is not confirming this; we’re sure it’s a Rebranded Phison E18. With eight channels, Phison’s PCI-Express 4.0 controller is manufactured on TSMC’s 12 nm node and features five Arm Cortex R5 CPU cores.  That SK Hynix chip is 2 GB of DRAM cache, stores the mapping tables.

And that’s Kioxia / Toshiba NAND right there, Toshiba 112-Layer 3D TLC BiCS5 TH58LKT2T45BA8H. 2TB for four chips, that’s 512GB per IC. 

Hardware & Software Used

We now begin the benchmark portion of this article, but first, let me show you our test system and the software we used:

Storage

• Kioxia Exceria Pro 2 TB SSD (2048GB/PCIe Gen 4)

Mainboard

• X570

Processor

• Ryzen 9 5950X 

Graphics Cards

• GeForce RTX 2080 Ti

Memory

• 16 GB (2x 8 GB) DDR4 3600 MHz

Power Supply Unit

• 750 Watt

Monitor

• Dell 3007WFP – up to 2560×1600

OS related Software

• Windows 10 64-bit

Software benchmark suite

• PCMark Vantage HDD test 1 through 8

• PCMark 8

• Anvil’s Storage Utilities

• ATTO Disk benchmark v2.4

• SiSoft Sandra Storage Benchmark

• HDTach

• HDTune Pro

• AS SSD Benchmark

Atto Disk Benchmark

One of the finest tools available to measure storage performance is ATTO. I love it to death as it is so reliable and produces such accurate results. The great thing about ATTO is that we can test with predefined block sizes. So we can test with a 32 MB sequence of 4 KB files, yet also 32 MB in 1 MB files. This gives us an excellent scope of overall performance with small and large files.

SSD Performance Application Traces

We now emulate real-world performance over several Windows applications like excel, adobe, and more.  Here we have three sets of results to share with us. Due to its excellent average access time result, we decided to include the benchmark in the test suite. It is quite significant.

SSD Performance Gaming traces

A lot of storage tests suffer from the flaw that they employ incorrect, synthetic workloads in order to estimate performance under ideal circumstances. In practice, the results of these tests are difficult to relate to practical and everyday requirements, which is why this Storage Benchmark is designed to measure real-world gaming performance. The input and output processes that make up storage activity are referred to as storage activity. The ability to record these actions while the storage device is in the middle of a task has been demonstrated. Traces are recordings that have been made.  This Benchmark analyzes traces recorded from popular games and gaming-related activities in order to assess real-world gaming performance;

• Loading Battlefield V from launch to the main menu.

• Loading Call of Duty: Black Ops 4 from launch to the main menu.

• Loading Overwatch from launch to the main menu.

• Recording a 1080p gameplay video at 60 FPS with OBS (Open Broadcaster Software) while playing Overwatch.

• Installing The Outer Worlds from the Epic Games Launcher.

• Saving game progress in The Outer Worlds.

• Copying the Steam folder for Counter-Strike: Global Offensive from an external SSD to the system drive.

SSD Performance Linear Read and Write performance analysis

In this chapter of the review, we will completely fill the SSD with data and observe what happens to the write performance as a result of doing so. So, for example, if the SSD has a capacity of 2TB, we’ll read and then write continually. So we can see how the SSD would react when subjected to large amounts of read and write activities at the same time.

As a solution to the sluggish read/write performance of TLC SSDs, manufacturers have added an SLC cache, sometimes known as a pSLC cache, into their products (pseudo). The amount of space available is determined by the capacity of the SSD. If there is still some free SLC cache space available, the SSD’s read/write performance is comparable to that of SLC NAND, as long as the space has not been totally utilized. When the SLC cache is completely depleted, the firmware (FW) initiates background garbage collection in order to clear off the remaining space. Additionally, while garbage collection is taking place, data is still being sent to the drive, which has an adverse effect on its overall performance. With this test, we can examine what happens when the cache fills up and the write speed falls back to the lowest feasible value.

We can see that the read performance remains constant across the full storage capacity (we’re reading 50% / 1 TB here); we average out at ~6000 MB/sec on average throughout the entire linear reading process.

Linear write testing then; When it comes to writing performance on any NAND storage device, the infamous Achilles heel comes into play – if the pSLC buffer runs out of space, and once that happens you’re back to bare TLC write performance. We peak at 6071 MB/sec after ~35% (700GB) percent written on the SSD, writes ~drop, and average around 600 MB/s. 

SSD Performance Video Capture

If there is one workload really stressful for any storage unit, it has to be video capturing. In this test, we simulate a 5K capture workload and test + log continuous writing towards the storage unit. The test is 64 GB in size and will expose any TLC/QLC write holes an SSD has, up to 64 GB, of course. Basically, we’re sending captures 5K frames towards storage and have it write/read as fast as it can. 

This 64GB file read test (readback of fame content) shows great results. You can see the peak read performance logged at ~6 GB/sec.

Writing (capturing video data) then, we write 64GB. Any TLC/QLC write holes (the moment when your SSD runs out of buffers) will be exposed here. And yeah, you can again see that a too-small pSLC buffer does not hinder the write performance. It can keep up quite well, though. 

HD Tune Pro (IOPS)

HD Tune Pro is a hard disk utility and benchmarking utility for Windows. We want to focus on IOPS perf in the dedicated test for that. IOPS (Input/Output Operations Per Second, pronounced eye-ops) is a performance measurement used to characterize computer storage devices like hard disk drives (HDD), solid-state drives (SSD), and storage area networks (SAN). IOPS numbers published by storage device manufacturers do not relate to real-world application performance, it does, however, create a metric of how many operations per second the storage unit is capable of. 

SSD Performance Crystal DiskMark

It is a disk benchmark application that assesses the performance of sequential and random read/write operations of varying sizes on any storage medium. CrystalDiskMark is a freeware disk benchmark utility. It is handy for evaluating the speed of different storage devices, both portable and local storage. When it comes to reads and writes, CrystalDiskMark can measure sequential reads and writes as well as random reads and writes of 512 KB, 4 KB, 4 KB (Queue Depth = 32) size. It also supports different types of test data (Random, 0 Fill, 1 Fill), has basic theme support, and supports multilingual input. You can give it a shot for free because it is available for download. The SSD is returning some pretty good results to us in terms of performance. Simply compare the read and write performance of the other drives.

Kioxia Exceria PRO 2TB Gen 4×4 

MSI M480 PLAY 2TB Gen 4×4 

Sabrent Rocket 4 PLUS 2TB Gen 4×4 (2022)

Sabrent Rocket 4 PLUS 2TB Gen 4×4

MSI M450 1TB Gen 4×4 

TeamGroup Cardea A440 PS5 PRO 4TB Gen 4×4 

Kingston KC3000 2TB Gen 4×4 

Corsair MP600 PRO XT 2TB Gen 4×4 

Corsair MP600 CORE 2TB Gen 4×4 

Silicon Power US 70 1 TB Gen 4×4 

Corsair MP400 4TB Gen3 x3

TeamGroup MP33 PRO 1TB Gen3 x4 

Samsung 980 Pro 1TB Gen4 x4 

Teamgroup Z440 1GB Gen4 x4

HP SSD 900EX PRO 1 TB

ADATA Falcon 1GB Gen3 x4

Sabrent Rocket 1TB Gen3 x4

Teamgroup C440 1GB Gen4 x4

Kioxia Exceria PLUS 2TB

Kioxia Exceria 1TB

Sabrent Rocket Q 1TB 

AS SSD Benchmark

Included in our benchmark suite is this nice little German application called AS SSD Benchmark. This test gives an extensive result set. The test is popular because it is so user-friendly and trustworthy in its results, so just for reference we always show the results, here we go.

Compression Benchmark

This test creates test patterns on the target drive which are random and vary in the level of compression used in the test data. This ranges from 0% compressible to 100% compressible. Basically, for this chart, you want a straight line as high as possible here (no compression performance loss) based on the highest performance in MB/sec.

HD Tach

HD Tach is a physical performance hard drive test that uses a special kernel mode VXD to get maximum accuracy by bypassing the file system.

Anvil’s Storage Utilities

Anvil’s Storage Utilities is a powerful tool that was designed in order to provide you with a simple way to evaluate the read and write performance of your Solid State Drive or Hard Disk Drive. The benchmark tool helps you monitor and check the response time of your unit as well as view the system information collected using Windows Management Instrumentation (WMI).

You can download this software here and try it out for yourself.  

Temperature Measurements

We monitor the SSD temperature during our test benchmarks stress tests with a high-stress workload. The peak stress temperature is listed below, we test as-is and that means if a heatsink is installed, we’ll show the value based on that. 

Above the temperatures reported from the sensors, this is with the heatsink installed. Below is a thermal image with the SSD writing at max perf. . Meaning seat this one wonder a heatsink if you have planned massive workloads.  

Final Words & Conclusion

Kioxia’s Exceria Pro is the company’s first high-end M.2 NVMe SSD with PCI-Express 4.0 capability. Paired with a rebranded Phison controller PS5018-E18-based device it just provides incredible performance. It really is fast, but not as fast as some of the many that compete at this level. Again let me reiterate, that Kioxia does not confirm the origin of the controller, but even the PCB layout seems to have been used from Phisons reference design. That said, in this conclusion, we are going to talk a little about relativity, though, as breaking that 7 GB/sec marker isn’t necessarily going to make your gaming PC a lot faster, and that is a simple truth. The good news is that Kioxia is not charging exorbitant prices for this outstanding device. Yes, this 7 GB/sec NVMe SSD comes at a high price, but at roughly 15 cents per GB, I am not unhappy whatsoever. This is a high-performance SSD with TLC writing, ample of endurance, and a 5-year guarantee. You’d be hard-pressed not to like this product at that price, wouldn’t you?

Do we really need 7000MB/sec storage units?

Um no? This is a premium performance product, often synthetically measured, and you’d need serious workloads to get the best out of it. Your PC isn’t going to boot faster as your OS is the bottleneck, your PC games might load a fraction of a second faster, your application load up just as quickly as an NVMe SSD with reads/writes in the 2 GB/sec marker. Guys, this is the honest truth. However, the same folk that purchases a GeForce RTX 3090 or Radeon RX 6900 XT combined with some sort Core i9 or Ryzen 9 series processor, it’s for those guys where that last sniff of performance matters, whether that is realistic or not, I’ll leave open to discussion.  In retrospect, however, we do have new technologies coming up like DirectStorage. This will allow the graphics card to load textures directly from the SSD bypassing the processor, freeing up processor cycles for other tasks, and speeding up texture load times. In this way, if they have a fast M.2 disk, they will be in the game in less than 5 seconds even on large maps, a negligible time compared to the loading times we are used to today. That technology will be released in the Windows 11 timeframe.

Endurance

Kioxia offers 400 TBW and 800 TBW (Terabytes Written) for each model, coupled with an MTBF of up to 1,600,000 Hours (Mean Time Between Failures). We talked so much about this in the past already, endurance, the number of times NAND cells can be written before they burst and shatter into small pieces (well, they just die and are mapped out, any data present on that cell is written to a healthy one). Bigger volume sizes mean more NAND cells; more NAND cells thus increase endurance. For the 2TB model, you’ll get a rated 800 TBW; the 1 TB model marks 700 TB written. So how long does a 800 TWB storage unit last before NAND flash cells go the way of the dodo? Well, if you are a really extreme user, you might be writing 50 GB per day (normal users likely won’t even write that per week), but based on that value, 50GB x 365days= 18.25 TB per year written. So that’s over 43 years of usage, half that for the 1TB SSD. And again, writing 50 GB per day is a very enthusiastic value. 

Thermals

The controller is not fitted with a heatsink or thermal sticker, the controller as such gets incredibly hot, resulting in thermal throttling under an extreme workload. Keep this SSD under a motherboard or 3rd party heatsink, and this will not be an issue.

Performance

The Kioxia Exceria Pro is a very fast product. Overall read performance increases from 3GB to well over 6.5 GB/sec. However, trace testing revealed that performance was somewhat lower, and I’ll have a guess as to why. The more NAND layers there are, it appears that access times become slower. We’re talking milli-fractions of a second here. The earlier generation, having fewer stacks, clocked in at (at most) 41 microseconds. It’s a pattern we’ve been observing a while now, the number increases as more additional layers are added; latency.

Concluding

The Kioxia Exceria Pro is an M.2 2280 NVMe SSD that utilizes Phison’s fastest E18 controller (unconfirmed by Kioxia) with newer firmware and 112 -Layer TLC 3D NAND flash to support capacities of up to 4TB. The SSD is 80.4 x 24 x 10.7mm in size and features a PCIe Gen 4 NVMe interface for compatibility with PCs. There are two storage capacities available: 1TB and 2TB, both come with a 2GB cache. Realistically the biggest performance benefit is deriving from the new B47 series Phison firmware, but small file performance has greatly increased with the latest updates. We do believe you will not notice the real-world effect and difference between a 3GB/sec and a 7GB/sec SSD anytime soon, but DirectStorage is getting supported, and that’s where it will matter. If an application loads in a fraction of a second, it will be quicker in that fraction of a second. Other than that somewhat personal remark, all lights are green, TLC, high endurance and warranty. Please do seat the SSD under a mobo-supplied heatsink. It’s running extremely hot without it, with a heatsink which prevents thermal throttling. Of course, to get the best out of it, you’ll need a PCIe Gen 4 infrastructure, and at the time of writing, that means a compatible Ryzen processor on, say, a B550 or X570 chipset-based platform. In March 2021, Intel will start with PCIe gen 4.0 support as well. The Exceria Pro 2 TB would cost €299 / $280 sans tax. The SSD is priced comparable to other high-end PCI-Express 4.0 drives at that price point, if not somewhat more. When the workloads match, the unit will make you shiver with performance, and when they are not, you will be limited to high-end NVMe performance (select workloads). NVMe protocol version 1.4; this SSD is one of the first to support the new NVMe 1.4 standard. You do not need to install any additional drivers; only ensure that your operating system is up to date. After that, install and format the SSD, and you’re good to go. Kioxia warranties the storage unit for a period of five years / or the claimed TBW value. It’s uncertain if you’ll ever require this level of performance or perceive a real-world difference between generation 3.0 and generation 4.0 devices. The good news is that as performance improves, both the low-end and mainstream markets will evolve. Consider for a moment that 3 GB/sec NVMe SSDs are fast becoming the industry standard and are presently readily accessible. Keep in mind that this SSD needs a heatsink of some sort to prevent throttling. The temperatures will not damage the product itself but will lower performance as it needs a bit of cooling. Most motherboards have excellent heatsinks already available. If you desire it, this is an enticing deal supported by a recognized brand; it has exceptional performance, a nice warranty, and high-quality components with a big name behind it all.