Training Throughput on Qwen3-30B-A3B workload
280,000Throughput (tokens/s)PithTrain
Evaluation Results
| Method | Links | |
|---|---|---|
| PithTrainHardware=4×8-H100, Parallelism=PP4,DP1,CP1,EP8, SeqLen=4096, Precision=BF162026.05 | 280,000 | |
| Megatron-LMHardware=4×8-H100, Parallelism=PP4,DP1,CP1,EP8, SeqLen=4096, Precision=BF162026.05 | 264,100 | |
| PithTrainHardware=1×8-B200, Parallelism=PP2,DP1,CP1,EP4, SeqLen=8192, Precision=FP82026.05 | 134,500 | |
| Megatron-LMHardware=2×8-H100, Parallelism=PP2,DP1,CP1,EP8, SeqLen=2048, Precision=BF162026.05 | 126,700 | |
| PithTrainHardware=2×8-H100, Parallelism=PP2,DP1,CP1,EP8, SeqLen=2048, Precision=BF162026.05 | 124,900 | |
| Megatron-LMHardware=1×8-B200, Parallelism=PP2,DP1,CP1,EP4, SeqLen=8192, Precision=FP82026.05 | 106,200 | |
| TorchTitanHardware=2×8-H100, Parallelism=PP2,DP1,CP1,EP8, SeqLen=2048, Precision=BF162026.05 | 90,500 |