List your company

Multi-Query Attention on T4 GPU Synthetic Performance Benchmark

19.46Performance (TFLOPS)DeepSeek-V3 + Ours
15.154416.272217.3918.5078Jun 14, 2025

Evaluation results

MethodLinks
DeepSeek-V3 + OursHead Dimension=128, Sequence Length=16k2025.0619.46
flash-attn v1Head Dimension=64, Sequence Length=1k2025.0615.32