List your company

LLM Training Throughput on Llama3-8B (4k SeqLen, train)

4,626.94Training Throughput (samples/sec)Megatron-LM
3,941.85044,119.71024,297.574,475.4298May 1, 2026

Evaluation results

MethodLinks
Megatron-LM2026.054,626.94
ZeRO-12026.054,261.65
AGoQ2026.053,968.2