List your company

LLM Training Throughput on Llama3-8B (2k SeqLen, train)

2,862.22Training Throughput (samples/sec)Megatron-LM
2,120.2842,312.9022,505.522,698.138May 1, 2026

Evaluation results

MethodLinks
Megatron-LM2026.052,862.22
ZeRO-12026.052,498.13
AGoQ2026.052,148.82