Almost a year after outlining their first roadmap for PCIe 4.0 SSDs, Samsung's first two models are in mass production: the PM1733 and PM1735 high-end datacenter SSDs. Details about these new models have been slow to come out, but Samsung is now talking about three major improvements they bring over earlier SSDs in addition to the raw performance increases enabled by PCIe 4.0. The list of improvements includes fail-in-place (FIP) technology to boost reliability of drives, SSD virtualization technology to guarantee consistent performance for VDI and similar use cases, as well as V-NAND machine learning technology to predict and verify characteristics of NAND cells.


Samsung’s fail-in-place (FIP) technology promises to allow the SSD to robustly handle hardware failures that would otherwise be fatal to the SSD, up to the failure of an entire NAND die. For the highest-capacity 30.72TB PM1733, the drive can keep running more or less normally even with the loss of any one of its 512 NAND flash dies. The drive will scan for corrupted or lost data, reconstruct it and relocate it to a still-working flash chip, and continue to operate with high throughput and QoS. In essence, this is like a RAID-5/6 array running in degraded mode instead of the whole array going offline. It's still wise to eventually replace a SSD after it suffers such severe malfunction, but Samsung's FIP technology means that replacement can be done at the operator's convenience instead of the problem causing immediate downtime.

The addition of fail-in-place doesn't change the fact that the PM1733 and PM1735 have write endurance ratings of 1 and 3 drive writes per day, respectively. The overall lifespan is still comparable to the previous generation of drives, but the chance of a premature death due to causes other than normal NAND wear has been greatly reduced.


Next up, Samsung has added virtualization technology to the PM1733 and PM1735 SSDs. Samsung has implemented the optional NVMe virtualization features based on Single-Root I/O Virtualization (SR-IOV), allowing a single NVMe SSD controller to provide numerous virtual controllers (up to 64 in the case of Samsung's drives). Each virtual controller can be assigned to a different VM running on the host system, and provide storage to that VM with no CPU overhead—the same as if the entire drive had been assigned to a single VM with PCIe passthrough. Storage capacity on each SSD can be flexibly allocated to different namespaces that can in turn be attached to the relevant virtual controller.

Machine Learning

The third technology introduced by Samsung is V-NAND machine learning. The company does not disclose precise details about how they are making use of machine learning, but only says that it is used to predict and analyze characteristics of flash cells, including by detecting variations among circuit patterns. With 3D NAND, it is increasingly difficult to get by with one size fits all strategies for cell programming, reading and error correction. Even tracking the P/E cycles each block has been through isn't enough; there can be significant variation between layers near the top and bottom of the 3D stack, and from one die to another. Samsung is hardly alone in turning to machine learning strategies to tackle these complexities. The new capability will ensure consistent performance and improved reliability of today’s drives powered by TLC V-NAND, but its importance will grow dramatically in the case of QLC V-NAND-based drives.

The first drives that can take advantage of the new features are already shipping to interested parties. The PM1733 and PM1735 are based on a common hardware platform. The PM1733 is rated for 1 DWPD and offers capacities up to 30.72 TB, while the PM1735 has more overprovisioning and lower usable capacities to reach 3 DWPD. Both models are available in either U.2 or PCIe add-in card form factors. The U.2 form factor gives a few more capacity options, while the add-in card versions have a PCIe 4.0 x8 interface to enable 25% higher sequential read performance (for other workloads, PCIe 4.0 x4 is fast enough to not be the bottleneck).


Related Reading:

Source: Samsung

Comments Locked


View All Comments

  • DanNeely - Friday, September 20, 2019 - link

    Fail in place seems like it should be capable of being expanded to handle multiple die failures at the expense of gradual performance degradation instead. 512 dies should be enough to afford a raid/ecc option that can afford to lose more than 1 die at a time. And since flash dies - unlike disks - can't be replaced to rebuild the array after a die fails instead of running in degraded integrity indefinitely, by reducing the amount of spare area available by however many dies are lost.

    Decline of protection vs decline of performance should be a configurable option; but outside of cache roles or highly distributed systems preventing the drive from bricking itself and causing a temporary system outage at the cost of a slight reduction in write performance seems like it would generally be the better trade off.
  • DyneCorp - Saturday, September 21, 2019 - link

    Are you THE Dan Neely?
  • DanNeely - Sunday, September 22, 2019 - link

    I am a Dan Neely but It's a common name, you'll need to define what you mean by 'The Dan Neely'. If you're talking about any of the ones that show up on page 1 of Google results for my name; no those are all some of my doppelgangers not me.
  • alpha754293 - Monday, September 23, 2019 - link

    "(for other workloads, PCIe 4.0 x4 is fast enough to not be the bottleneck)."

    In my experience, the interface CAN be the bottleneck, but only from a theorectical perspective.

    I'm already running into the situation/scenario where between the host/client communication subsystem and the storage subsystem, there aren't enough PCIe 3.0/4.0 lanes because all of it can be consumed/accounted for in order to keep the data feed coming.

    I'm still working on verifying a testing methodology to benchmark and test the interaction between high speed network communications and high speed data storage subsystems in order to maximise and improve the performance of both.
  • tmr.sofyan - Wednesday, March 11, 2020 - link

    flagships are great all the time,

Log in

Don't have an account? Sign up now