2020 OCP Virtual Summit

Deep learning training workloads call for a balanced system design with an optimal balance of compute, acceleration and memory.  In this talk we discuss how we leverage system design knobs across compute, memory, storage & networking, coupled with the systems & hardware management tools and software to develop an optimized solution for customers training needs from small scale workloads running on a few cards to large scale complex workloads spanning multiple chassis & racks. We will showcase how the flexibility of using different fabric solutions along with the Inter Chip Link enables different topologies and highly efficient training system configurations to be built. We will share how we design to strike a balance between system performance, software complexity and total cost of ownership.