Because of all the cabling kerfuffle involved with the 2.5-inch drive and the limits of many M.2 implementations, we figure most users will stick with the expansion card drive. Intel might have had the same thought when the company sent us that model to test.
Even before we installed the Intel NVMe drivers (Windows 8.1 has its own), we saw 1.3GBps reading and 900MBps sequential writing (4MB) scores in CrystalDiskMark 3. With the Intel drivers, that jumped to 1.6GBps reading and 1.4GBps writing. Small 4KB writes were rated at 229MBps for individual writes and significantly, 832MBps with the queue depth set at 32. Part of NVMe's magic is its multiple queues.
Because Intel was throwing around numbers like 2.5GBps, however, we switched to Iometer. Using Intel's recommended 128KB file size, we saw just about that. Cool beans. Using a beta version of CrystalDiskMark, it displayed around 2.7GBps.
Things were looking 2.5GBps' worth of fantastic. Then we ditched the server scenarios and headed back down the client-desktop road, where things slowed down again.
Most users only pay attention to disk performance when they copy a large amount of data from one place to another. So to back up the artificial benchmarks we copy a large 20GB file, and a 20GB file and folder mix, back and forth. Because NVMe and SSDs are so fast, we generally do this with a RAM disk to eliminate the secondary drive as a bottleneck.
The best we saw in this ad-hoc testing when reading from the 750 series was 1.5GBps with the large 20GB archive and 1.2GBps with the files and folders. We copied manually, using XCOPY from the command line, and called the Windows file and folder copying routines from a Visual Basic script. Same basic results. Not bad, but not quite as fast as we were expecting.
We moved on to PC Mark to see how adding a 750 series would impact overall system performance. Here the 750 series scored quite a bit better than the system running on a single SATA SSD, but it actually lost to the RAID 0 SATA SSD array (two Samsung 840 Pros) that shipped with our Core i7-5960X/X99, 32GB Polywell X9900E4. The score difference of 3,493 to 3,503 is statistically meaningless, but it does show that you can get a lot of bang for your buck from SATA RAID.
What exactly was throttling the raw sequential throughput — Windows, the RAM disk, the NVMe drivers, or simply the drive itself — we were unable to ascertain in time for this review. We suspect that part of the story is that, while you'll see large gains over SATA simply because NVMe is using PCIe, most client usage is relatively straight-line and won't take advantage of everything NVMe has to offer. Multi-track audio and video editing, transcoding, and the like may see greater boosts than simply copying a single file.
Sign up for CIO Asia eNewsletters.