The DPURVDRML is a high-performance general RNN processing engine optimized for the Versal ACAP AI Core series. Versal devices can provide superior performance or wattage over conventional FPGAs, CPUs, and GPUs. The DPURVDRML is comprised of AI Engines and PL.
The GEMM operation with a precision of int8 is deployed on the 5x8 AI Engine array. Each AI Core performs matrix-matrix multiplication of size 32x64x32. The 40-core kernel calculates GEMM of size 32x320x256. The output of GEMM is quantized to int16.
The Misc modules in the following image are composed of different modules to support different types of operations, including element-wise multiplication, addition, sigmoid, tanh, and max. Intermediate data should be represented with the precision of int16.
