AIBench: An Industry Standard Internet Service AI Benchmark Suite
The user manual of AIBench: http://www.benchcouncil.org/AIBench/files/AIBench-User-Manual.pdf
The micro benchmarks of AIBench includes 12 workloads. The workloads are implemented not only based on mainstream deep learning frameworks like TensorFlow, but also based on traditional programming model like Pthreads, to conduct an apple-to-apple comparison. The datasets used for micro benchmarks are Cifar and ImageNet.
The 12 micro benchmarks are as follows.
- No. DC-AI-M1: Convolution,
- No. DC-AI-M2: Fully Connected
- No. DC-AI-M3: Relu
- No. DC-AI-M4: Sigmoid
- No. DC-AI-M5: Tanh
- No. DC-AI-M6: MaxPooling
- No. DC-AI-M7: AvgPooling
- No. DC-AI-M8: CosineNorm
- No. DC-AI-M9: BatchNorm
- No. DC-AI-M10: Dropout
- No. DC-AI-M11: Element-wise multiply
- No. DC-AI-M12: Softmax