As big data and artificial intelligence become increasingly standard infrastructure in today's society, cloud data center hardware and chips are evolving to meet the needs of big data and artificial intelligence. An important feature of Big Data and AI is the need for massive amounts of data, either generated by Internet users or by servers after analysis, and while these massive amounts of data provide core support for Big Data and AI, they also place new demands on data storage.

In traditional storage architectures, this data is stored in SSDs, which are installed in specific storage servers. When the server needs data, it needs to communicate with the storage server first, and the storage server then goes to read the data from the SSD into memory, and the CPU goes to do a series of underlying tasks (such as filtering, encryption and decryption, etc.) to finish reading or writing the data from or to the SSD. the whole process requires the storage server CPU to handle a large number of such underlying tasks, thus becoming the energy of the SSD read and write The CPU in the storage server also becomes a significant part of the cost when large-scale deployment is required.

To solve such storage efficiency problems, compute storage is an important direction. With compute storage, the SSD hardware will become intelligent by integrating compute and processing chips, thus giving a lot of repetitive underlying tasks directly to these SSD integrated compute and processing chips, thus reducing the need for CPUs in the storage server - which can be used to handle Some higher-level tasks. In intelligent SSDs, many of the underlying tasks do not require a particularly high-end CPU, so by integrating a more basic CPU in the SSD, the amount of CPU computing in the storage server can be reduced; at the same time, many of the underlying tasks in intelligent SSDs can even be implemented with solidified digital logic, which further increases the efficiency of computing and processing, as the CPU is more efficient for such specialized tasks. The processing efficiency, especially the energy efficiency ratio, is often very low.

The idea of compute storage has been around for several years, but its true maturity and entry into the mainstream requires that the storage industry as a whole define a standard that can be widely used to ensure that different companies can have a uniform interface to their intelligent SSD storage. The Storage Networking Industry Association (SNIA) has established a dedicated working group for this purpose, including 58 member companies (including Samsung, Magnesium, Huawei, Wave and other storage and system vendors around the world).

At the end of August this year, after four years of discussion, the SNIA Compute Storage Task Force published the first version of the Compute Storage Architecture and Programming Model standard, which became a milestone event for intelligent SSDs, because with this standard completed by the industry's mainstream companies, the industry can truly implement the corresponding hardware and software design and application in mainstream products.

The technical path to SSD intelligence

As the concept of intelligent SSDs is gradually accepted and standardized by the industry, we expect more and more intelligent SSD-related products in the future, and the demand for related processing chips in intelligent storage products will gradually increase. Here we analyze the chip technology used in the currently announced intelligent SSD products and predict the future direction of development.

Samsung is one of the storage giants that has invested the most in intelligent SSDs. Samsung's relevant product is the SmartSSD Computational Storage Device (CSD). The first generation of CSDs was released in November 2020 with Xilinx FPGAs, while the second generation of CSDs was just released in July this year with the same Xilinx (AMD) processing chips, but updated with Versal FPGAs. Adaptive SoC (which includes FPGAs and ARM cores). According to Samsung, its second-generation CSD products can do data filtering, data compression and format conversion locally, which can reduce the energy consumption of storage servers by 70% and improve the CPU load of storage servers by up to 97%.

The FPGA solution used by Samsung focuses more on flexibility, as FPGAs are programmable and therefore have better compatibility for user software interfaces and related tasks. According to Samsung, its second-generation CSDs run mainly on customer-written software and IP, so it makes sense to use FPGAs to ensure maximum flexibility and compatibility.

In addition to Samsung, startup ScaleFlux has also released its line of intelligent SSDs. Last November, ScaleFlux released its third generation of intelligent SSDs, the CSD 3000, whose technology differs from Samsung's in that it uses an ASIC SoC solution rather than an FPGA solution. The integrated SoC in its SSD products includes an octa-core ARM, hardware acceleration engine (including data compression, data encryption, data matching and acceleration of hash algorithms), and flash memory and PCIe interface control.

Comparing the two technology solutions using FPGA and SoC, the FPGA technology solution is flexible and the NRE cost is lower when the shipments are low, but the cost may become its bottleneck after high volume shipments; while the SoC is the opposite, less flexible than FPGA and the NRE cost is high, but the cost will have an advantage after high volume shipments. We believe that as smart SSDs are gradually recognized by the mainstream, the relevant standards will be fixed slowly on the one hand, and the volume of shipments will gradually become larger on the other hand, making SoC solutions more attractive in the future. However, FPGA solutions are not useless, and we believe that different customers will have customized needs for their SSD products in terms of computing and software. In this regard, we believe that the final SoC product will also need to be programmable and configurable to meet customer needs for customization.

Competitive landscape and future market for compute storage

As big data and artificial intelligence continue to evolve, we believe that intelligent SSDs and compute storage will continue to be used more and more. As mentioned earlier, the biggest players in compute storage include traditional storage giants (e.g. Samsung) and startups (e.g. ScaleFlux), and more companies are bound to enter this market in the future.

In terms of the competitive landscape, Samsung is currently the largest investment in compute storage among the mainstream storage companies, and has already shipped products, so it is expected to accumulate the most relevant experience and gain a leading position in the future. Other mainstream vendors are currently in R&D status, such as SK-Hynix, which jointly developed a CSD for accelerated data retrieval with Los Alamos National Laboratory earlier this year, a milestone in R&D but not yet in shipping status. Of course, there are opportunities for startups to gain market share as product development is generally at an early stage.