International Business Machines Corporation

Erich Jurg Hauptli

Daniel Mark Dreps

Prasanna Jayaraman

A method, a system and a computer program product for supporting a reconfigurable hardware network topology including graphics processor units (GPU) and central processing unit (CPU) interconnectivity. The re-configurability of the network is based on data bandwidth and latency requirements of running workloads obtained by running training workload sequences upon the configured topology. For network re-configurability, a user enabled to directly connect high-speed cable links between CPU/GPU connectors and between GPU/GPU connectors. Further included in a configured network topology are switches configurable to provide one or more PCIe high-speed side-band links, as well as a high speed multiplexor, that can switch the network topology. The method prepares a system map of the hardware network topology, and generates a GPU weightage pattern based on data bandwidth and latencies of the configured GPUs such that GPUs are assigned to workloads based on its weightage to optimize workload performance.

Reconfigurable network infrastructure

A method, a system and a computer program product for reconfiguring hardware network topology including graphics processor units (GPU) and central processing unit (CPU) interconnectivity on or across compute nodes of a rack-mount server. The re-configurability is based on detected thermal throttling or thermal hot spots when running workloads. For network re-configurability, a user can directly connect high-speed cable links between CPU/GPU connectors and between GPU/GPU connectors on a same PCB compute node, or across two PCB compute nodes as suggested by a control processor to avoid thermal and power hotspots when running the workload. The method recommends and generates a system map of the hardware network topology known to avoid/mitigate thermal throttling, and instructs a configuration of CPUs and GPUs such that GPUs are assigned to workloads at locations for mitigating thermal throttling based on detected thermal hot spots and power hot spots to optimize workload performance.

Growing community of inventors

Cedar Park, TX, United States of America

Erich Jurg Hauptli