Wireless enabled Inter-Chiplet Communication in DNN Hardware Accelerators

Inter-chiplet communication is a fundamental bottleneck in scale-out Homogeneous Multi-Chip-Module-based Hardware Accelerators (HMCMHAs). This paper focuses on the problem of many-to-many communication traffic generated when dispatching output feature map tiles among chiplets. Such traffic has a strong impact on the latency and energy metrics of the HMCMHAs as it exposes the limitations of the existing wire-based Network-on-Package (NoP). This paper investigates augmenting the existing NoP with emerging wireless in-package communication links. The intrinsic single-hop and broadcast-capable technology is exploited to tackle the many-to-many communication traffic in question. We show that the proposed wireless-enabled NoP can significantly improve the latency and energy of Deep Neural Network (DNN) inference on HMCMHAs.