Incident Library

Organisation

Samsung SDS

Samsung SDS · samsung-sds

1 case · 2026

Root cause Software defect

Component NCCL PyTorch Training data loader

Contributing factor Monitoring blind spot Gap in automated tests

Symptom NCCL timeout Training job failure

Cases

IL-0480 Epoch-end NCCL deadlock hung a 16-GPU B300 fine-tuning run for three hours while GPU utilization read 100%Samsung SDS · 2026
Root cause Software defect