By Jamon Bowen, Director of Product Marketing, Computational Storage
Xilinx Data Center Group
Computational storage continues to gain attention as the performance benefits of processing data near where the data resides has been well established. Industry momentum is building and organizations like the Storage Networking Industry Association, or SNIA, are helping to define the how, what, where, and why of computational storage through the Compute, Memory, and Storage Initiative. One of the computational storage types SNIA has helped to identify is a Computational Storage Processer (CSP), where a compute function can be done on the same bus as peer storage devices. This enables the data to be fetched for processing and returned to the storage devices without the expensive interrupt handling and cache memory pollution that would occur if the CPU was responsible for this data movement.
One of our partners, Bittware, just announced a Xilinx FPGA-based CSP leveraging the M.2 form-factor. This FPGA-powered CSP was designed to meet the new Open Compute M.2 Accelerator Module standard. Within the Open Compute Project (OCP) community, infrastructure has been created to enable small accelerators, or datacenter SSDs, using the storage M.2 form-factor (with some enhancements ) to be connected to servers and peer storage devices. Together with Bittware, we’re extending CSPs into the OCP ecosystem where Glacier point M.2 carriers can share these CSPs and SSDs amongst Yosemite modular server nodes (OCP specification for these hardware solutions are available on their server wiki: https://www.opencompute.org/wiki/Server/Working).
By leveraging the flexibility of our adaptable devices, Bittware is showcasing workloads diversity from storage compression and encryption from Eideticom to ML inference recommendation model acceleration from Myrtle.ai.
CSPs are a way to realize the acceleration of data-intensive functions, and when combined with the flexibility of our adaptive computing devices, they enable both a custom acceleration and a standard platform. There is always a question on how much acceleration is needed, and with the goal of CSPs to be on the same bus as the peer storage devices, there is a need for the form-factor of a CSP to fit in systems at the right place. For example, our Alveo™ U50 datacenter acceleration card can take on a CSP persona when placed in an AMD server, where all of the SSDs are peers to the Alveo U50 due to AMD’s PCIe architecture. In other platforms, a different form-factor could be needed.
Bittware’s new Xilinx-based CSP is the latest example of a novel form factor and we’re excited to help enable this new addition to the CSP landscape and the OCP hardware community. It will bring hyperscalers and cloud companies improved performance density and energy efficiency for their machine learning platforms.