基于可变精度脉动阵列结构的卷积神经网络加速器

秦婷,李钊,闫玉婷,杜亚飞

Convolutional neural network accelerator based on variable-precision systolic-array structure

Ting QIN,Zhao LI,Yuting YAN,Yafei DU

表 3 CIFAR-100数据集上不同方法的Top-1准确率对比

Tab.3 Top-1 accuracy comparison of different methods on CIFAR-100 dataset

模型 方法 量化精度 Top-1/% ResNet20 Baseline FP32 74.46 PACT INT4 65.56 LSQ INT4 71.37 LIMPQ 4MP 73.95 本研究 INT4/8 74.01 ResNet56 Baseline FP32 74.10 PACT INT4 70.35 LSQ INT4 73.56 LIMPQ 4MP 73.80 本研究 INT4/8 74.06