List your company

Multi-Query Attention (MQA) Performance on NVIDIA A100 GPU

211.4TFLOPScuDNN
154.304169.127183.95198.773Jun 14, 2025

Evaluation results

MethodLinks
cuDNNHead Dimension=128, Sequence Length=16k, Causal Mask=false2025.06211.4—
DeepSeek-V3 + OursHead Dimension=64, Sequence Length=512, Causal Mask=false2025.06156.55.36