Samsung 950 PRO M.2 NVM Express SSD
The latest piece of hardware I've been playing around with at Phoronix is Samsung's V-NAND SSD 950 PRO M.2 NVM Express SSD. Assuming you are running a modern Linux distribution, this M.2 PCI-E NVMe SSD can offer blazing fast performance.
The Samsung 950 PRO M.2 NVMe internal SSD (MZ-V5P256BW) is slated for maximum sequential reads of up to 2200 MBps and maximum sequential writes of up to 900 MBps. Samsung rates the 4KB random reads at up to 270,000 IOPS and the 4KB random writes at up to 85,000 IOPS. This NVMe driver is rated at 1.5 million hours MTBF, has a 512MB cache, and uses Samsung's UBX controller. Pricing on this NVM Express drive is around $180 USD for the 256GB unit and around $320 for the 512GB version.
Samsung lists the operating system support as Windows 8.1 and Windows 10 for offering native NVMe support, but of course I've only been playing with it under Linux. The original NVMe Linux support has been present since the Linux 3.3 kernel while since then there's been improvements to benefit NVMe storage like the blk-multiqueue code. Long story short, if you are running a Linux distribution released in the past year or so, you should be in good shape for NVMe on the software side.
All of my Samsung 950 PRO M.2 testing has been done so far on the MSI C236A Workstation motherboard with Intel Xeon E3 v5 Skylake CPU. For those curious about the performance potential of the Samsung 950 PRO NVMe SSD, I've posted some Ubuntu 16.04 benchmarks of it on the following pages.
I compared the Samsung 950 PRO performance to a few SATA 3.0 SSDs I have on hand: Samsung 850 EVO, OCZ Vertex 3, and Intel SSD 530. Unfortunately with rarely receiving review samples of storage devices at Phoronix, this comparison isn't against any other NVMe drives. However, thanks to the Phoronix Test Suite for benchmarking, you can easily run benchmarks of your own system(s) side-by-side to the results in this article for seeing how the disk performance compares to your own hardware. EXT4 on Linux 4.6.0 was used for all of the benchmarking in this article.