A proposed solution to improve reliability by containing the impact of PCIe Uncorrected Errors

Presented by Anil Agrawal (Meta) | Gada Badeer (Meta) Meta's next generation of AI/ML platform called "Grand Teton Training" uses a complex hierarcy of PCIe devices including GPUs, Switches, NICs, and NVMe drivers and require various RAS features to improve the system reliability. In this presentation, we would share the key learnings as we developed this platform and propose an optimized solution to contain any risk of data corruption due to PCIe uncorrected errors.