Follow Datanami:
June 9, 2022

Samsung to Ship Next-Generation Smart SSD This Year

Agam Shah

Samsung will ship a new version of its “smart” SSD for data centers that’s more than just a storage device — it can be a CPU and accelerator when needed.

The second-generation SmartSSD computational storage device (CSD) will ship to customers for evaluation later this year, said Yang Seok Ki, vice president for the memory solutions lab at Samsung Electronics, during a presentation last month at SNIA’s Persistent Memory and Computational Sstorage Summit.

The SmartSSD can store multiple terabytes of data and has on-device processing capabilities with an embedded CPU and an accelerator.

Servers can offload processing and acceleration to the storage device’s processors to reduce the strain on CPUs and save network bandwidth. The applications could be storage-specific tasks such as compression and data warehousing.

“CSD can provide scalable low-power processing with a low-bandwidth footprint at the rack level,” Yang said, adding “our next-generation device is focusing on developing a platform that enables various collaborations and technologies.”

CSDs could use FPGAs, RISC-V chips, AI accelerators, or custom chips called ASICs. The CSD could be used to filter or compress data, meaning it could free up I/O, bandwidth, and cache and improve network utilization. The on-device processing capabilities ensure only relevant data leave storage.

Samsung’s 2nd-generation CSD combines storage and processing, and will ship later this year (Image courtesy Samsung)

The processors can run background applications that are low priority and not required to run in real time. Accelerators can focus on foreground applications with strong real-time requirements, but within a specific power budget.

Samsung released its first SmartSSD CSD product in 2020, with a 4TB SSD based on the company’s fifth generation V-NAND technology, and a Xilinx FPGA on which customers can install their own applications. The FPGA accelerator talks directly with the SSD without going through the host CPU.

The first-generation SmartSSD, which is based on PCI-Gen3 protocol, has a sequential read rate of 3,300 MB/sec, write rate of 2,000 MB/sec, and random read 100,000 IOPS and random write speed of 800,000 IOPS.

In a video transcoding application, Samsung measured an 80% boost in performance by breaking up processing between the CPU and FPGA on the SmartSSD, and that reduced server CPU utilization from 93% to 70%. It also sped up cyber forensics, data warehousing, and compression applications.

The next-generation SmartSSD compute module will include an embedded CPU, and a separate computing core where customers can put it in their own software stack depending on the operation. Those processing modules will connect to an SSD controller, which will be linked to NAND storage modules.

Tech specs for Samsung’s second-gen CSD (Source: Samsung)

The upcoming device is more software-defined, and customers will be able to add their own intellectual property to the custom processors, Yang said.

Software developers can expand device functions by adding custom functions to the processors. Samsung later this year will release a software development toolkit that includes a library that is compatible with the SNIA’s computational storage APIs and specifications.

The next-gen SmartSSD device will also have system-level security features that include advanced encryption standards.

Yang wrapped up the need for computational device storages like SmartSSD as a way to improve the energy efficiency of data centers.

“We need a new computing paradigm to reduce emissions in data centers. In the current situation where large amounts of data are generated and lots of computing power is required, data-centric computing with computational storage devices can be a promising alternative,” Yang said.

A few trends have helped make servers more efficient, including new manufacturing technologies, which has increased the power efficiency of chips. About 92% of servers are virtualized, and 80% of the data centers are cloud based or cloud-like, which has helped reduce the idle time of servers, Yang said.

Servers use about 50% of the data center power consumption and reducing that is important to meet the targets, Yang said, citing numbers from a paper published in Science Magazine in 2020.

The growth in storage-drive efficiencies and densities has enabled a 25-fold increase in storage capacity with only a threefold increase in global storage energy use, the paper stated.

Yang stated that offloading processing to storage devices could improve overall energy efficiency of data centers and reduce greenhouse gas emissions.

“We must do better for the next-generation,” he said.

Related Items:

AWS Adds a Little More Nitro to Its SSDs

Micron Ships New Data Center SSDs, Teases DDR5