Paper Title: Accelerating NVM to GPU Transfer Rates Using PCIe Fabrics and Logical Volumes

Paper Abstract: A critical performance metric in every AI/ML workload is the rate at which training or inference data can be transferred to GPU memory. This topic will discuss how GPU direct transfers, PCIe fabric technologies and logical (RAID) storage volumes can be applied to substantially increase the rate at which GPUs can access these datasets from NVM storage. GPU direct transfers allow data to be directly transferred between NVM devices direct and GPU memory, without being copied to host memory. PCIe fabrics can route these peer-to-peer transfers directly across the fabric, decreasing root port congestion and surpassing CPU performance bottlenecks. Employing logical storage volumes delivers maximum instantaneous bandwidth to the NVM storage. The combination of these technologies offers blazing fast transfer rates of files direct from NVMe storage into GPU memory.

Paper Author: Vincent Hache, Principal Applications Engineer, Microchip Technology

Author Bio: Vincent is the Firmware Principle Applications Engineer for Microsemi's Switchtec PCIe Switches, and the software architect and product owner of ChipLink, Microsemi's advanced cross-platform debug and diagnostics tool. His current focus is enabling PCIe fabrics using the Switchtec PCIe advanced fabric switches to provide a flexible fabric interconnect for shared pools of GPUs and/or NVMe SSDs to accelerate AI and Machine Learning applications. Vincent has been providing architectural and engineering support for Tier 1 customers’ storage, server, and data center designs in the areas of NVMe and GPU interconnect, and Serial Attached SCSI (SAS) and PCIe infrastructure over the last 7 years. Prior to working in the data center solutions industry, Vincent worked in systems and software engineering for radar surveillance systems.