This is a simulator for evaluating the inference performance of In-DRAM CNN accelerators like DRISA, ATRIA, SCOPE, and LACC
git clone https://github.com/Sairam954/PIM_ACC_SIM.git
python main.py
The accelerator configuration can be provided in config.py file. The configuration dictionary looks like below:
DRISA_1T1C_NOR = [{
NAME: "DRISA_1T1C_NOR", # Name of the accelerator, this name will used for logging performance results in accelerator_performance.csv file
PE: 32768, # Number of Processing Elements in the accelerators
TRC: 8, # ns
MUL_CYCLE: 200, # Latency to perform multiplication in ns
ADD_CYCLE: 21, # Latency to perform Addition in ns
DATA_MOVE_LATENCY: 8, # Latency corresponding to the data movement in ns
DATAFLOW: "WS", # Execution dataflow, currently framework can only map using Weight Stationary (WS) dataflow. However, it can be extended to map Input Stationary and Output Stationary dataflows
MAC_ENERGY: 6630, # Energy consumed to perform a MAC operation in pJ
BATCH_SIZE: 1 # The input batch size
}]PIM_ACC_SIM
├── acclerator_performance.csv # Results of Simulations are stored in this file. Simulator captures metrics like Total_Run_Time, Total_Data_Movement, Total_MAC_Energy, and BottleNeck_Ratio
├── Atria_PAPER_results_modified_V13.xlsx # Excel sheet containing the latency, energy, and area values corresponding to different accelerators, these values are used to create configuration dictionaries in config.py
├── Config.py # File is used to declare the accelerator configurations used for evaluation
├── ConvolutionToGEMM.png # Illustration of how convolution operations are converted to GEMM operation before mapping onto accelerator
├── main.py # Selecting the accelerator and model for evaluation are set in this file. Currently, the complete logic of mapping and calculating the performance metrics happens in this file.
│
├───CNNModels #Folder contains various CNN models available for performing simulations.
│ ├── DenseNet121.csv
│ ├── GoogLeNet.csv
│ ├── Inception_V3.csv
│ ├── MobileNet_V2.csv
│ ├── ResNet50.csv
│ ├── ShuffleNet_V2.csv
│ ├── VGG16.csv
│ ├── VGG19.csv