Read and Write Benchmarks

As we mentioned, we are going to post screenshots of the results.  In the future, we will have comparison data so you can see how drives lineup.

CrystalDiskMark

As you read the results keep in mind that the quoted read speed is 3500MB/s and the quoted write speed is quoted at 3000MB/s but will depend on the application according to XPG’s performance table. 

CrystalDiskMark shows very good read results of what we’d expect for this drive.  It is coming in at 3477MB/s which is right at that 3500MB/s quoted read speed.  These are really good results and prove the quoted speed.  The write speed, however, is a bit slower than what the quoted speed is.  Here we are seeing a maximum of 2288MB/s write speed.  This is by means a slow number but is just shy of the quoted speed.  It’s still great NVMe performance though, and as we’ve learned the write speed will be different depending on the application and load types.  This test overall proves the SSD is fast on read and write for a TLC drive.    

ATTO Disk Benchmark

Next is the popular ATTO Disk Benchmark, we are using the default settings.  Drive performance is really excellent on the read speeds according to this benchmark.  We are seeing very consistent read speeds down the line at 3240MB/s, which is just shy of 3500MB/s but still good, and more importantly consistent. 

The write speeds are again a bit slower.  We did see a maximum of 2140MB/s write, which is close to the result of CrystalDiskMark.  But, as the file size grows this write speed slows down to 1970MB/s, so just under the 2000MB/s mark.  It’s still a good performance, but not as high as we’d expect for this drive.  There are SSDs out there most likely with higher sustained write speeds, but again we have to keep in mind this is TLC, so for TLC, it is good.

Aida64 Disk Benchmark

For our last test we are running Aida64’s Read Test Suite.  In the future we may also run the write tests, but those take hours to run.  On the read tests here we are seeing around 3097-3137MB/s read speeds.  This is lower than CrystalDiskMark, still above 3000MB/s at least, but lower than the quoted 3500MB/s.  With differences that small it’d be impossible to tell the difference in real-world scenarios, but this does show that the SSD is just a bit slower than quoted, but for the money not bad.

Recent Posts

Brent Justice

Brent Justice has been reviewing computer components for 20+ years, educated in the art and method of the computer hardware review he brings experience, knowledge, and hands-on testing with a gamer oriented...

Join the Conversation

9 Comments

  1. Very nice to see you guys starting to review these devices. Great work Brent!

    I’m starting to fill my home lab with these devices, trying to determine if they are suitable for use as storage for a SQL Server. Since I won’t be able to hardware raid these with the host, I’m looking at simple Mirroring done within the OS for redundancy, and seeing what kind of performance I’m gonna get with 250GBs worth of SQL data.

  2. Some things that stood out. By no small means instead of by small means. ;)

    Also when you were talking access times you were using MB/s. Isn’t that a throughput number and not an access time? I was expecting some .02 MS access times.

    Otherwise good article. I would be willing to capture some performance data on my older NVME 3.0 Corsair for comparison if it would help.

  3. Only 128MB of DRAM on this drive (64M x 16b) which is par for the course on the higher-end Realtek drives, which includes the SX8100, S40G (RGB w/heatsink), and SX8800 (single-sided). Their lower-end offerings are DRAM-less with HMB.

    DRAM on consumer drives is typically used for caching metadata, particularly addressing, to translate between logical and physical locations. Such mapping can be expensive in terms of memory required, for example 32-bit (4-byte) addressing for 4KiB requires 1GB per 1TB of NAND/flash. DRAM can be used for other things but it’s not really used as a write cache (you do have coalescing of subpage or <4K writes, but this is more commonly done in SRAM). So the hottest data stays in DRAM, although all controllers have some amount of SRAM of which part is used for metadata.

    SLC caching, on the other hand, is fully a write cache. The goal is the empty it to TLC as soon as possible. ADATA tends to have large, dynamic SLC caches which has advantages and disadvantages. It’s more flexible but can add wear and is less consistent e.g. when the drive is fuller or after sustained writes. Most commonly the "on box" speeds you see for drives are using SLC program times for the seq. write speed and TLC read times for the seq. read (on the order of 200-300µs and 80µs, respectively, with 16KiB pages).

    Realtek controllers tend to run hot although I’ve heard a firmware update on these can help with that. Still, it is a legacy of Realtek controllers in general. They’re not quite as mature as the competition’s offerings (SMI, Phison, etc).

    Despite the SX8100 being rated for up to 3000 MB/s, we see that it’s closer to 2 GB/s or so, although some of that is from lower queue depths. There’s been speculation by some reviewers (W1zzard at TechPowerUp) that this uses pMLC (MLC caching) rather than pSLC (SLC). This is theoretically possible as the program time of good flash is in the 450-500µs range in MLC mode:

    x = (1/0.000475)(64)(0.015625), x = ~2100 MB/s. 64 = 64-way interleaving (8-channel, 4 dies/channel, 2 planes/die) and 0.015625 is the page size in MiB. Although there’s no way it can hit the rated 3000 with current controllers in my opinion, and further even Samsung’s 96L TLC in MLC mode is only 500µs. It’s possible it has both modes but I haven’t tested these drives myself. More likely it just doesn’t have great low queue depth sequential writes.

    Other than that for this review, I consider these drives halfway between budget options and higher-end consumer NVMe drives. You have reduced DRAM, performance isn’t quite as consistent, the controller/firmware is okay but needs optimization, it’s actually not a bad option at 2TB though.

    1. Hey my dude can you help troubleshooting my ssd? If anyone knows anything about this, please let me know. I bought the Sx8100 2 months ago and could use it without any problems on my laptop (wich uses raid for local storage), after installing and playing games as many other everyday tasks it used to be just fine, after some time i got crashes on games (precisely Cod warzone) wich would just tell me the game was broken (problem with files) and even though i tried openning the folders on the ssd, some folders were just not accessible. I have tried using the ADATA program to find any issues and when it is on this "dead" state, it wont recognize properly anything at all. It gets fixed after a reboot but it happens freqntly, I am afraid it might be some sort of compatibility issue cause if it was hardware it should come and go randomly, right?
  4. I’m starting to fill my home lab with these devices, trying to determine if they are suitable for use as storage for a SQL Server. Since I won’t be able to hardware raid these with the host, I’m looking at simple Mirroring done within the OS for redundancy, and seeing what kind of performance I’m gonna get with 250GBs worth of SQL data.

    Well, these drives are designed around a very large, dynamic SLC cache which makes them oriented at consumer workloads. They won’t be as consistent, especially when fuller or after sustained writes, and the endurance will potentially be lower (although that’s generally not a significant factor for this segment). They also have reduced DRAM; 128MB at 2TB is pretty diminutive. If you’re doing a lot of writes, especially small (file) writes, and extra-especially with a fuller drive, this is not ideal. Reads will come from the native flash (TLC) so is a different story, although MLC (e.g. 970 Pro) would be faster there. Generally I consider TLC-based drives with static SLC to be the best "consumer" (retail) drives for workspace/workstation use, that would include the WD SN750 and 970 EVO Plus for example (the EVO Plus has a hybrid cache, but I digress). More powerful controllers are also ideal. Etc.

  5. Extra note on this:

    Looks like the upcoming ADATA Falcon will be a DRAM-less version of this drive with a heatspreader, that is RTS5762DL (HMB).

  6. I’d recommend getting ahold of Adata. This sounds like a drive losing its ‘state’ and may be a firmware issue, or they may need to replace your drive.
Leave a comment