Results: Write Endurance Testing
Over the past few years, we've added a lot of write endurance testing to our benchmark suite. The rise of read-focused enterprise drives made those metrics even more important. Fortunately, during the last 12 months, testing enterprise SSDs standardized partly due to the introduction of JEDEC's JESD218A write endurance testing standard. Instead of issuing vague ratings, we now see companies specifying their drives to JESD218A, which uses the JESD219A enterprise workload to quantify endurance. The good news from all of that is that this closely matches the types of workloads we use in our Enterprise Workload Performance tests.
Previously, our write endurance testing consisted of large block sequential writes, which is completely different from the JEDEC standard. Still, it gives us valuable information about P/E cycles and allows us to compare SSDs at a NAND level, reducing the effects of flash controller optimizations and over-provisioning. Write amplification is also reduced so that we don't incur any additional NAND writes beyond what we intend.
For this test, we use the 240 GB model and monitor SMART attribute 177, which is the wear leveling count. It's a normalized value meant to show a percentage of life remaining on the drive. We took multiple percentage points off of the drive and compiled the results. We then extrapolated the number out to a full 100% usage; the following numbers are what we ended up with:
|Endurance RatingSequential Workload, QD=1, 8 MB||Samsung 845DC EVO||Intel SSD DC S3500|
|NAND Type||Samsung 19 nm 3-bit MLC||Intel 20 nm MLC|
|Raw NAND Capacity||256 GB||528 GB|
|IDEMA Capacity (User Accessible)||240 GB||480 GB|
|P/E Cycles Observed (IDEMA)||3643||3651|
|P/E Cycles Observed (Raw)||3181||3097|
|Host Writes per 1% of MWI||7.95 TB||15.97 TB|
As you can see, the 845DC EVO actually does get more P/E cycles (according to SMART) with Samsung's 3-bit MLC than we observed from Intel's SSD DC S3500.
Does that sound too good to be true? Let's change things up a but, then. Instead of using large block sequential writes, we'll go with 4 KB random writes instead, which should get closer to the JEDEC specification and apply more stress to the NAND/controller. Also, in this test we'll use the 960 GB model to see if the difference in over-provisioning matters.
The 687 TBW result ends up being mostly the same. Still, you're getting almost 15% more than Samsung's 600 TBW specification. That doesn't come anywhere close to eMLC flash or some of the optimized MLC-based solutions from SanDisk and Micron, but considering the price and target segment, it is more than adequate.
I'm also not taking into account that many review sites, including this one, have already pushed their consumer 840 EVO drives far beyond the point at which the wear leveling count hits zero. We hope our testing helps convince some of the doubters, but we can appreciate skepticism amongst a high-end audience.