Lightning Talk: Efficiency and Programmability of DNN Accelerators and GPUs

While GPUs have provided high computational power for AI applications, their power efficiency may not be satisfied for applications requiring a high-speed response within a low power budget. Specialized deep neural network (DNN) accelerators have been developed to address this challenge. However, there are still limitations and concerns surrounding the adoption of DNN accelerators as a viable choice for commercial products. This lightning talk aims to highlight the importance of fast and domain-specific DNN accelerators and discuss the additional features they should possess besides speed and power efficiency. By addressing these points, the talk will shed light on the necessity of specialized DNN accelerators and explore potential areas for improvement to meet the growing demands of efficient and high-performance AI computations.