EAGLE: Exploiting Essential Address in Both Weight and Activation to Accelerate CNN Computing

Efficient quantization techniques can compress Convolutional Neural Networks (CNNs) with less bit-width while maintaining the accuracy on large extent. However, the quantized CNN models hardly boost the computation performance of CNN accelerators with the conventional bit-parallel Multiply-Accumulate (MAC) operations. Previous works proposed a shifting-based bit-serial operation, which can be called as Shift-Accumulate (SAC) operation, to take advantage of the reduced bit-width. However, it is also found that there are many invalid computations in both MAC and SAC operation, caused by zero bits in activations and weights, which are not optimized. To fully exploit the computations in CNN models, we proposed a Essen-tial Address only GAC based Low-latency Efficient (EAGLE) architecture that can further accelerate the CNN computation through bypassing zero bits computation in the activations and weights. An essential address is adopted to encode the nonzero bits in activations and weights in this architecture. Furthermore, to support the essential address-only computations, Generate-Accumulate (GAC), an operation which produces partial sums with essential addresses, is implemented. The architecture is implemented with a TSMC 28nm CMOS technology. Based on the results, if scaled in a 65nm technology, the EAGLE only requires 63.6% area and 43.1% power consumption compare to that of the Pragmatic. The EAGLE reaches an average speedup of $ 2.08\times$ and $ 1.43\times$ on six CNN models over the Stripe and Pragmatic at a similar frequency, respectively. It also improves energy efficiency by $ 3.69\times$ on average over the DaDianNao baseline.