Before delivering on that BuzzFeed bait, I should take a moment to introduce the competition. We've tested the SM951 against a bunch of PCIe SSDs, including its XP941 predecessor. The results also include Plextor's M6e, an early M.2 drive with a much slower dual-lane Gen2 interface.
Intel's 750 Series matches the SM951's Gen3 interface, but it comes on a much larger expansion card with a beefy heatsink. This product of datacenter trickle-down is derived from the server-grade DC P3700, which we've included strictly for reference. The P3700 sells for roughly two bucks per gig and devotes a much larger percentage of flash to overprovisioned area. It's a very different class of PCIe SSD.
Speaking of a different class, the results also cover a stack of the latest SATA SSDs, plus the old Intel X25-M. This collection should give us broader context for the SM951's performance.
IOMeter — Sequential and random performance
IOMeter fuels much of our new storage suite, including our sequential and random I/O tests. These tests are run across the full extent of the drive at two queue depths. The QD1 tests simulate a single thread, while the QD4 results emulate a more demanding desktop workload. (87% of the requests in our old DriveBench 2.0 trace of real-world desktop activity have a queue depth of four or less.) Clicking the buttons below the graphs switches between the different queue depths.
Our sequential tests use a relatively large 128KB block size.
The SM951 beats the 750 Series with reads—and quite comfortably at QD1. It's substantially slower in the write speed tests, though. Samsung's new hotness even slips behind the old XP941 when slinging sequential writes at QD4.
"New hotness" is more than just a reference to Men in Black; the data suggest the SM951 is throttling to prevent overheating. The graphs above show the median result from a trio of three-minute averages, but IOMeter also logs transfer rates for every second of each run. A sampling of that data is presented below, along with results from a separate test that cooled the drive with direct airflow from a 120-mm fan.
Yikes. Although the uncooled config reaches 2000MB/s in the read speed test, it only holds on for about 25 seconds before plunging below 100MB/s. The slowdown lasts for less than ten seconds, after which the drive ramps up for another spurt at top speed.
Sequential writes are evidently more problematic, forcing the SM951 to spend more time in the valleys than it does at the peaks. Similar oscillations are also evident with sequential writes at QD1—but not with reads at that queue depth. Meanwhile, the air-cooled config maintains top speed over the duration of each test.
The uncooled XP941 experiences similar slowdowns, just to a lesser degree. It's the only other SSD we've seen exhibit this behavior.
For what it's worth, our IOMeter tests run consecutively, with no breaks between them. The drive is also filled with a test file immediately before the benchmark begins, effectively conditioning it with a sequential write workload. The overall pattern of peaks and valleys is consistent from one run to the next, but the duration of the initial burst is different in each run, likely due to whatever residual heat is left over from the previous test.
Next, we'll turn our attention to performance with 4KB random I/O. We've reported average response times rather than raw throughput, which we think makes sense in the context of system responsiveness.
Impressively, the SM951 has the quickest read response times of the bunch, beating even Intel's best datacenter SSD. The Samsung drive's read response times are consistent across each test run, as well.
Oscillations appear again in the random write tests, where the SM951 is less competitive with the Intel PCIe drives. To be fair, though, the differences between the SSDs amount to just fractions of a millisecond. Also, all the drives suffer a decline in random write performance over time. We explore that slowdown more closely on the following page.