Whither Computational Storage?

This week, Pliops announced $100m in additional funding, while at FMS 2022, we heard NGD Systems was not doing so well. In the few short years that we’ve known about Computational Storage, how are the successful companies managing to maintain momentum?

Background

Modern computing systems are built around the concept of centralised processing, memory, and peripherals. However, look under the covers, and you’ll see that components like network switches and storage also have the same framework in place. Solid-state disks and hard drives have embedded controllers that perform the translation between the host’s logical block-level view of storage and the myriad layers of complexity involved in safely storing data on disk platters or NAND flash.

The Practicality of In-Situ Processing

As we’ve reported in the past, a plethora of new startups has emerged to address the computational storage opportunity, including NGD Systems, ScaleFlux, Bittware, Pliops and potentially, IBM. These vendors are either offloading compute into a storage device or using an add-in-card (AIC) to accelerate or otherwise process data.

(Note: there are more vendors on the market today, however in this brief post we’re covering those we’ve already reviewed and had conversations with).

NGD Systems

The NGD Systems model uses embedded compute on a solid-state disk to perform analysis tasks on the content being stored, what can be referred to as “in-situ processing”. Theoretically, a computational SSD could analyse data in real-time as an edge device records video or still images to it. The result of the analysis could be object identification or something more complex like number-plate recognition.

#117 – Introduction to Computational Storage with NGD Systems (Storage Unpacked)

At first glance, this concept seems pretty cool, however there are some obvious initial issues with this technology. First, the device must be formatted with a local, visible (to the controller) file system. These devices would be impossible to implement as RAID-protected storage as each device would hold only a fragment of data. In an edge scenario, this may work, but not in the enterprise. Second, to make the data visible (if not encrypted), the drive would need to be trusted with the file system encryption keys. It isn’t at all clear what protocols are in place to wipe failed drives, so key management becomes a challenge.

Third, the delivery method for active code is via the host. A drive assumes the host is trusted. Obvious issues arise around data injection, rogue devices, and other security issues. So far, I’ve seen nothing to explain how these challenges will be resolved.

ScaleFlux

ScaleFlux has a slightly different approach to computational storage, with a range of AIC and U.2 devices that implement onboard compression and optimisation. This results in a drive with greater capacity, performance and endurance than a typical drive using the same volume of NAND flash. We spoke to ScaleFlux around 18 months ago when the most recent product on offer was the CSD2000 series. At FMS 2022, ScaleFlux announced the CSD3000 series, which is based on an ARM SoC design. The new product range supports PCIe 4.0 and the new E1 format (in addition to U2 and AIC).

#194 – ScaleFlux & Computational Storage Devices (Storage Unpacked)

ScaleFlux is effectively a cost efficiency solution. Customers get more capacity and performance for the same money, with the capability to push products harder (greater endurance). Most important, the CSD devices are drop-in compatible NVMe drives, so require no additional software or configuration. This makes the technology appealing to hyper-scalers and other large-scale enterprise vendors that are focused on TCO.

Pliops

Pliops just announced an additional $100 million with Series D funding. At FMS 2022, we caught up with Pliops Global VP of Products & Marketing, Tony Afshary, to discuss the latest developments in the company’s technology.

#180 – SmartNICs – Pliops Storage Processor (Storage Unpacked)

The Pliops XDP (Extreme Data Processor) is an AIC that sits in a server with NVMe drives, exposing either block storage or a Key/Value API to the host. XDP manages the physical storage directly, effectively implementing what initially looks like a RAID card. However, this analogy is a gross simplification of the capabilities XDP provides. We’ll look at the technology in more detail in a separate post, however, like ScaleFlux, Pliops’ technology is also a cost, reliability, and efficiency play.

IBM

Finally, in our brief discussion, there’s IBM. IBM has been developing a computational storage SSD called the FlashCore Module (FCM) with the technology acquired from the acquisition of TMS in 2012. FlashCore Modules implement data reduction through compression, allowing IBM to offer the largest capacity NVMe SSD drive in the market today.

#236 – A FlashCore Module Primer with IBM Fellow & Storage CTO Andy Walls (Storage Unpacked)
#235 – FlashCore Futures with IBM CMO Scott Baker (Storage Unpacked)
IBM FlashSystem Review – Part 1 – Hardware

As we learned from IBM Fellow Andy Walls, there are also additional FCM features to make use of QLC flash, and in the future, these drives could implement a “hinting” technology that enables the host to identify challenges such as ransomware or just improve performance.

The Architect’s View®

This rundown describes just a few of the computational storage products and vendors in the market today. The obvious difference between the winners and losers in this market is that successful products are focused on efficiency and cost reduction. This outcome should come as no surprise, where the largest customer for this kind of technology will be hyper-scalers and large-scale enterprises. There’s also justification for HPE and AI/analytics applications. Today’s cost/efficiency computational storage devices don’t expose any standard APIs to the host. This may be an area where hyper-scalers would like more control. This is something for discussion in another post.

We simply don’t see a market (yet) for smart SSDs that work at the data level. This is because this type of implementation is so much more complex and introduces questions of security and management. Until (and unless) standards are developed to fix these problems, data-aware computational storage devices are likely to wither on the vine.