HDTach

HD Tach will test the sequential read, random access and interface burst speeds of your attached storage device (hard drive, flash drive, removable drive, etc). All drive technologies such as SCSI, IDE/ATA, 1394, USB, SATA and RAID are supported. Test results from HD Tach can be used to confirm manufacturer specs, analyze your system for proper performance, and compare your performance with others. HD Tach is very easy to use, quick, and presents data in easy to read graphs, including the ability to compare two storage devices on screen at the same time for easy analysis.

Bursts are provided only for your review. SSD's don't cache the same way HDD's do (in many cases they don't cache reads at all), so burst testing typically results in figures that are lower than the sequential throughput figures, regardless of controller used.

HDTach feeds the tested drive a continuous string of small sequential requests. This is a single threaded operation, which means the SSD doesn't get to see what's coming next. The lower the QD=1 latency of the controller pipeline, the better the numbers we see from this test. It doesn't equate to real-world maximum throughput, but it does mean something for analysis, which is why we include these results. Here we can see how the various controllers handle this type of workload:

  • In the wrk, the Silicon Motion SM2246EN produces QD=1 sequential reads matching only the Samsung 850 Pro, perhaps the highest we've seen from any SATA SSD on this particular test. Writes were not as high, but still satisfactory, and close to the drive's rated speed.
  • Marvell (Crucial) sees limited read speed, as each IO request in series requires an LBA map table lookup, which is forced to occur in series, slowing the read speed overall. Not shown on the graph is that the read speeds were ~450 MB/sec on the first pass (before the drive was fully written).
  • The Samsung 850's triple core ARM controller, paired with their 32-layer VNAND perform with very low latency.
  • The Intel 730 uses an overclocked version of their own DC S3500/S3700 enterprise controller, which was purpose-built with low IO latency in mind, so high scores here are a given.
  • The low latency Indilinx M10 controller does well in QD=1 operations.

This is the sort of reason why we keep these benches around. Their output may be dated, but it still has meaning, and when interpreted correctly, can tell us things other benches have a hard time replicating.

« PreviousNext »