THE SQL Server Blog Spot on the Web

Welcome to - The SQL Server blog spot on the web Sign in | |
in Search

Joe Chang

Samsung SSD 960 PRO - 3500MB/s

Last month, Samsung announced that their 960 Pro, (PCI-E x4 gen3) M.2 NVMe SSD would available in October at 512GB, 1TB and 2TB capacities. This is a high-end PC and workstation product. There is a similar but separate product line (SM961?) for OEM customers. The 960 Pro had already been mentioned earlier this year. This week, all the hardware sights published product reviews, including: Anandtech, Tomshardware, TheSsdReview and others as well.

All the previous PCI-E x4 gen3 NVMe SSDs were rated between 2,000-2,500MB/s in large block read. The 960 Pro is rated for 3,500MB/s read. This is pretty much the maximum possible bandwidth for PCI-E x4 gen3. Each PCI-E gen3 lane is 8Gbit/s, but the realizable bandwidth is less. In earlier generation products, an upper bound of 800MB/s realizable per 8Gbit/s nominal signaling rate was typical.

Presumably there was a reason why every PCI-E x4 was in the 2000-2500MB/s bandwidth. It could be that these were 8-channel controllers and the NAND interface was 333-400MB/s. Even though 8 x 400MB/s = 3,200MB/s, it is expected that excess bandwidth is necessary on the downstream side. The could be other reasons as well, perhaps the DRAM for caching NAND meta-data. Intel had an 18-channel controller, which produced 2,400MB/s in the P750 line, and 2,800MB/s in the P3x00 line.

It is very annoying that the more vertical SSD OEMs decline to disclose meaning details of the controller in their SSD. The 960 PRO uses the new Polaris controller, for which few details beyond the Read/Write IOPS and bandwidth specifications are published. On TheSsdReview, the 960 PRO is shown to have the Polaris controller and DRAM in a single package. There are 4 NAND packages, each with 16 stacked 512Gbit die for the 2TB product.

I am guessing that the Polaris controller has 8-channels, at 8-bit per channel with each package connecting to 2 channels? But there are other possibilities, such as 4 x 16-bit channels.

The reason is that the previous generation NAND were 333 or 400MB/s. The most recent NAND from Micron is listed at 533MB/s. In order to support 3,500MB/s on the upstream interface, there needs to be more than that amount on the downstream side combined. While Samsung is the leader in V-NAND, it seems more likely that the 960 has 8 channels 8-bit (or 4 channels, 16-bit) at 533MT/s and not 4 channels, 8-bit 1066MT/s. Between the two likely scenarios, 8-channel 8-bit should support higher IOPS than 4 channel 16bit?

Both the 2TB and 1TB models have read specs of 440,000 IOPS while the 512GB model is 330,000 IOPS. It could be that the 1TB is 4 packages, 16 chips stacked with a 256Gb chips, and the 512GB model has 8 chips stacked. It is also possible that the 1TB has packages stacked 8 high with the same 512Gb, just that there is no reduction in IOPS, while the 512GB has 4 chip stacks?

I would think that 2 x 960 PROs in a single socket Xeon E3 quad-core would be a great match. My current test system is a single socket Xeon E5 10-core with four Intel 750 SSDs. I will probably not outfit another test system until the Xeon E5 v5 comes out next year, in which I may opt for the 960 PRO.

Oh yeah, the 960 PRO is not a enterprise product, so should probably not be used in critical production environments, but should be ok for test.
It is possible to buy a M.2 to PCI-E adapter, but I would like an adapter that can accommodate 2 M.2 devices in one PCI-E x8 slot, without the use of a PCI-E switch if possible.
HP has an adapter for 4 M.2 in a x16 that only works in HP systems

last year I bought a Dell XPS9550 with the then new Skylake processor. It came with a 512GB Samsung PM951, which was an older model rated for 1GB/s sequential, while the slightly more expensive SM951 was rated for 2GB/s. This drive died after 3 months. Dell sent a replacement, but I elected to go out and buy the 950 PRO myself, rated the same 2GB/s as the SM951, but perhaps same or not? The failure of one device is not sufficient to make an assessment, but I have dozens of OCZ Vertex and Vectors, and several Intel SSDs of various flavors, all running for several years now without failure.

Published Wednesday, October 19, 2016 6:48 PM by jchang

Comment Notification

If you would like to receive an email when updates are made to this post, please register here

Subscribe to this post's comments using RSS



OCZ No Go! said:

Every OCZ product I have ever owned has died within one year of operation.  Even with Toshiba buying OCZ line, I am forever reluctant to use them.

November 9, 2016 11:04 AM

jchang said:

could you be specific on which OCZ products, ex, Vertex/Vector, or the low end?

November 13, 2016 9:26 AM

Leave a Comment


About jchang

Reverse engineering the SQL Server Cost Based Optimizer (Query Optimizer), NUMA System Architecture, performance tools developer - SQL ExecStats, mucking with the data distribution statistics histogram - decoding STATS_STREAM, Parallel Execution plans, microprocessors, SSD, HDD, SAN, storage performance, performance modeling and prediction, database architecture, SQL Server engine

This Blog


Privacy Statement