Our Test Methods: Under each test condition, the SSDs tested here were installed as secondary volumes in our testbed, with a separate drive used for the OS and benchmark installations. Out testbed's motherboard was updated with the latest BIOS available at the time of publication and AHCI mode was enabled for the host drive. The SSDs were secure erased prior to testing (when applicable), and left blank without partitions for some tests, while others required them to be partitioned and formatted, as is the case with the ATTO and CrystalDiskMark tests. Windows firewall, automatic updates, and screen savers were all disabled before testing and Windows 10 Quiet Hours / Focus Assist was enabled to prevent any potential interruptions.
In all test runs, we rebooted the system, ensured all temp and prefetch data was purged, waited several minutes for drive activity to settle and for the system to reach an idle state before invoking a test. All of the drives featured here were tested with their own
NVMe drivers installed where possible / available, but the default Windows 10 NVMe driver was used when a proprietary driver was unavailable. Also note, we have completely revamped our test bed, so the numbers shown in this review aren’t comparable to previous articles. All of the drives here have also been updated to their latest firmware and drivers where applicable.
 |
| HotHardware Test System |
| Intel Core i9 Powered |
|
Processor -
Motherboard -
Video Card -
Memory -
Audio -
Storage -
|
Intel Core i9-9900K
Gigabyte Z390 Aorus Master
(Z390 Chipset, AHCI Enabled)
Intel HD 630
16GB G.SKILL DDR4-2666
Integrated on board
Corsair Force GT (OS Drive)
Samsung SSD 883 DCT (960GB)
Samsung SSD 883 DCT (1.92TB)
Samsung SSD 883 DCT (3.84TB)
Samsung SSD 983 DCT (960GB)
Samsung SSD 983 DCT (1.92TB)
Intel SSD DC4510 (2TB) |
OS -
Chipset Drivers -
DirectX -
Benchmarks - |
Windows 10 Pro x64 (1809, 17763.475)
Intel 10.1.1.45, iRST 17.2.0.1009
DirectX 12
IOMeter 1.1
HD Tune v5.70
ATTO v4.00.0f2
AS SSD
CrystalDiskMark v6.0.2 x64
SiSoftware SANDRA |
|
| IOMeter |
| I/O Subsystem Measurement Tool |
|
As we've noted in previous SSD articles, though IOMeter is clearly a well-respected industry standard drive benchmark, we're not completely comfortable with it for testing SSDs. The fact of the matter is, though our results with IOMeter appear to scale, it is debatable whether or not certain access patterns, as they are presented to and measured on an SSD, actually provide a valid example of real-world performance. The access patterns we tested may not reflect your particular workload, for example. That said, we do think IOMeter is a reliable gauge for relative available throughput, latency, and bandwidth with a given storage solution. In addition, there are certain higher-end workloads you can place on a drive with IOMeter, that you can't with most other storage benchmark tools available currently.
In the following tables, we're showing two sets of access patterns; a custom Workstation pattern, with an 8K transfer size, consisting of 80% reads (20% writes) and 80% random (20% sequential) access and a 4K access pattern with a 4K transfer size, comprised of 67% reads (33% writes) and 100% random access. Queue depths from 1 to 32 were tested...
As you can see, the SATA-based Samsung 883 DCT drives are not capable of reaching the same levels of performance as an NVMe-based drive like the 983 DCT, as the queue depth and demands placed on the drives increases. In relation to the Intel solution, the Samsung 983 DCT drives are competitive throughout, though
the Intel drive finishes ahead at the lower queue depths, before things cross over and the Samsung drives pull ahead in the fully-random 4K workload. With 8K / 80 / 80 test, however, the Intel drive pulls ahead at all stops, except QD8.
The actual bandwidth numbers obviously jibe with the number of peak IOs, so the lines in the charts look identical. What you'll note is that the SATA-based 883 DCT drives bump into the limits of their legacy interface once they hit QD8. The NVMe drives, however, are able to offer much more bandwidth and exceed the 1GB/s mark with the fully random 4K access pattern and about 1.7GB/s with the 8K access pattern that includes some sequential transfers in the mix.
Latency characteristics for the drives are inversely proportional to available bandwidth. Here, the NVMe-based drives offer the lowest latency throughout testing and remain tightly grouped overall. The Samsung 983 DCT drives have a slight latency advantage with the random 4K access pattern, but the tables turn in favor of Intel with the 8K test.
|
| AS SSD Compression Benchmark |
| Bring Your Translator: http://bit.ly/aRx11n |
|
Next up we ran the Compression Benchmark built-into AS SSD, an SSD specific benchmark being developed by Alex Intelligent Software. This test is interesting because it uses a mix of compressible and non-compressible data and outputs both Read and Write throughput of the drive. We only graphed a small fraction of the data (1% compressible, 50% compressible, and 100% compressible), but the trend is representative of the benchmark’s complete results.
The compressibility of the data being transferred across the drives we tested had little to no impact on performance. The NVMe-based drives from
Intel and
Samsung finished right on top of each other in the read test, and save for the lower-capacity 960GB 993 DCT, the same rang true in the write test as well. The SATA-based drives, however, once again bump into the limits of their legacy interface.
