List your company

Training Throughput on Llama3 8B (train)

2,320.47Throughput (128K SeqLen)Ulysses
1,125.72841,435.90171,746.0752,056.2483Feb 24, 2026

Evaluation results

MethodLinks
UlyssesHardware=8xH100s2026.022,320.471,503.8878.63475.33246.05162.41——
UPipeHardware=8xH100s2026.022,281.051,487.29867.17472.53246.07166.32125.5698.25
RingHardware=8xH100s2026.022,064.91,387.67841.05458.51237.99159.96——
Native PyTorchHardware=8xH100s2026.021,373.87845.99474.3249.85————
FPDTHardware=8xH100s2026.021,171.68884.75621.2382.42219.53153.48119.76—