KIOXIA EXCERIA PRO 2TB Review – BiCS5 has to compete with Micron’s 176-layer NAND, also in terms of price

1 - Einführung, Unboxing, Technik und Daten 2 - Teardown: Controller und NAND im Detail 3 - Dynamischer pSLC Cache im Detail 4 - Testsystem und synthetische Tests 5 - Anwendungsbenchmarks und Zusammenfassung

Test system and test preparation

I use the usual suspects like CrystalDiskMark and Atto to check the theoretical information from the specs. However, I do not make it easy even for these programs, because the SSDs were written to several times with about 66% of the storage space. Thus, these are not brand-new SSDs, but rather everyday goods that have already been worn down. Let’s see what remains of the theory in everyday life after wear and tear. The SSDs under test are located in the second NVMe slot of the motherboard and are not used as a system disk.

In addition, I use AJA as an everyday test to simulate the encoding of larger Ultra HD video streams and the storage test of SPECwpc, which contains a lot of real applications and it will be interesting to see what performance is left for the large workloads. However, I picked out the applications with the biggest differences and loads as examples here. The whole thing runs on my current small workstation with the Ryzen 9 5950X and the MSI MEG X570 Godlike along with 32 GB DDR4 3800.

I have also summarized the individual components of the test system in a table:

Sequential performance of used SSDs

Synthetics are a great way to really break out the big numbers once in a while. We will see later how well this works in reality with the real application benchmarks. Therefore, I start with the CrystalDiskMark and four different file sizes. The SSDs were no longer new at the time of the test (I always do these tests at the end for certain reasons) and I also had fill levels of plenty of 50% before the data was deleted several times. This certainly also explains why we are a bit short of the maximum values, but can still show impressive figures. That is certainly why the wording with the “up to”.

And at least here, thanks to adapted firmware, you still have the nose ahead of the SSD with the Micron NAND in a single scene with the BiCS5 when it comes to sequential reading of the large blocks alone. The rest is dominated by Micron.

KIOXIA EXCERIA PRO 2TB NVMe	MSI SPATIUM M480 2TB NVMe PLAY

You can see very well that the dynamic pSLC does exactly what it is supposed to, mind you with an empty (though not virgin) SSD. The nice thing about the 2 TB SSD is that there is a lot of space left and you should therefore never fill it more than 2/3 with data. A higher load does not affect reading, but the dynamic SLC will certainly reach its limits at some point during writing. And if you do it over and over again, switching the memory modules between the two methods will eventually become impossible as well.

ATTO also works very similarly, although I only work with two sizes here, which ends up being the same. And again, we see a similar picture: There is a tie in reading, but BiCS5 is unfortunately clearly defeated in writing. The focus of a gaming SSD is rather on reading. So it still fits quite well up to here.

KIOXIA EXCERIA PRO 2TB NVMe	MSI SPATIUM M480 2 TB NVMe PLAY

But what happens when you stream a video? For this purpose, the industry uses the AJA benchmark, which is in effect an interface between synthetic benchmarks and practical application. The EXCERIA PRO does not fluff here either, even though it clearly lags behind the SSD with Micron’s 176-layer NAND. Let’s first look at the writing process of the encoded video content and then compare the whole thing in the table with the competitor:

The minimum rate is similar, but the KIOXIA SSD lags a tad behind in both peak performance and average.

Reading is also quite fast, although you don’t reach the theoretically possible speed here either. Interestingly, both SSDs don’t differ much in pure streaming

We see that the comments made on the previous page about the dynamic pSLC cache and the behavior with the larger file blocks are completely true. Smaller file movements would be even faster if you leave out the overhead of the file system.