List your company

End-to-end attention latency measurement on RK3588S2

2.95Attention Latency (ms)IntAttention
-88.1304526.66231,141.4551,756.2477Nov 26, 2025

Evaluation results

MethodLinks
IntAttentionSequence Length=1K2025.112.95
Quant-OnlySequence Length=1K2025.115.96
FP16Sequence Length=1K2025.1110.96
IntAttentionSequence Length=2K2025.1111.29
FP32Sequence Length=1K2025.1112.57
Quant-OnlySequence Length=2K2025.1122.24
FP16Sequence Length=2K2025.1136.04
FP32Sequence Length=2K2025.1142.09
IntAttentionSequence Length=4K2025.1146.96
Quant-OnlySequence Length=4K2025.1184.85
FP16Sequence Length=4K2025.11134.64
FP32Sequence Length=4K2025.11160.11
IntAttentionSequence Length=8K2025.11200.94
Quant-OnlySequence Length=8K2025.11325.48
FP16Sequence Length=8K2025.11440.3
FP32Sequence Length=8K2025.11586.86
IntAttentionSequence Length=16K2025.11794.36
Quant-OnlySequence Length=16K2025.111,279.94
FP16Sequence Length=16K2025.111,677.49
FP32Sequence Length=16K2025.112,279.96