Generating Acceleration Systems for Next Generation Deep Learning Models
GeneSys is a programmable deep learning acceleration system generator. The core computation engines in GeneSys are a systolic array (for operations such as convolution) and SIMD engine (for operations such as ReLU and pooling). GeneSys is parametrizable, and it is possible to automatically generate hardware with different numbers of processing elements, bit-widths, on-chip buffer configurations, and memory bandwidths. The generated accelerator acts like a co-processor connected to the host via the PCIe bus. The target workloads for the accelerator are Computer Vision and Transformers.