The Computational Storage Processor (CSP) has been developed to meet the new Open Compute M.2 Accelerator Module standard. The module is intended to work in Glacier Point carrier cards for Yosemite servers. The feature-rich, dense servers are used by hyperscale and cloud companies who want to increase the performance density and energy-efficiency of machine learning platforms.
The 250-M2D offers a fully programmable Xilinx Kintex UltraScale+ FPGA that has been directly coupled to two banks of DDR4 memory. The FPGA can be programmed by customers for in-house capabilities, or provided as a ready-to-run pre-configured solution with application IP from Eideticom. The solutions can be purchased from BittWare’s parent company, Molex. Molex is a trusted provider to many of the world’s largest data centre customers. One version of the module has been specifically developed for recommendation models, and comes complete with software for easy integration into deep learning frameworks. This version is available direct from Myrtle.ai.
“Our continued investment in NVMe-based Computational Storage products, plus exciting collaborations with application experts such as Myrtle.ai, will help our customers innovate quicker and with lower risk,” said Craig Petrie, vice president marketing at BittWare. “FPGAs offer unique and valuable capabilities. It is essential that there is a supplier who can deliver and support higher-quality volume deployments. BittWare, as a part of Molex, is in the unique position to drive technology advancements while simultaneously delivering enterprise-class product.”
“When co-designing the hardware and software for our deep learning based recommendation model accelerator, SEAL, we selected BittWare, part of the Molex group of companies,” said Peter Baldwin, CEO at Myrtle.ai. “It was a clear choice to take what we’d designed to volume. BittWare had the technical expertise to deliver an enterprise-class product and, with Molex’s global logistics, there was an established channel to support volume hardware for hyperscaler and Tier 1 customers.”