[{"id":"IL-0478","title":{"en":"A failed GPU network adapter made NCCL silently reroute through GPU 0's adapter and slowed MoE training steps","ko":"GPU 한 개의 네트워크 어댑터가 고장 났는데 NCCL이 알림 없이 통신을 우회시켜 MoE 학습 스텝이 느려졌다"},"org":"Amazon","org_slug":"amazon","date":"2026-05-18","severity":"unknown","root_cause":"hardware-failure","url":"https://benchlet.com/incidentlib/app/cases/IL-0478","shared_components":["network-link"]},{"id":"IL-0586","title":{"en":"Four NIC-to-T0 links flapped in rapid succession and cut a 50K-GPU pretraining job's throughput about 25% for a minute","ko":"T0 스위치 트랜시버의 링크 네 개가 연달아 플래핑해 50K GPU 사전 학습 작업의 처리량이 1분 동안 약 25% 떨어졌다"},"org":"OpenAI","org_slug":"openai","date":"2026-05-05","severity":"unknown","root_cause":"hardware-failure","url":"https://benchlet.com/incidentlib/app/cases/IL-0586","shared_components":["network-link"]},{"id":"IL-0418","title":{"en":"Machine, ICI link and optical circuit switch failures interrupted training jobs on Google's TPU v4 supercomputers","ko":"TPU v4 슈퍼컴퓨터에서 장비, ICI 링크, 광 회로 스위치가 고장 나 학습 작업 약 1% 가 멈추거나 느려졌다"},"org":"Google","org_slug":"google","date":"2024-04-16","severity":"unknown","root_cause":"hardware-failure","url":"https://benchlet.com/incidentlib/app/cases/IL-0418","shared_components":["network-link"]}]