Energy-Efficient Reconfigurable XGBoost Inference Accelerator With Modular Unit Trees via Selective Node Execution and Data Movement

The extreme gradient boosting (XGBoost) has emerged as a powerful AI algorithm, achieving high accuracy and winning multiple Kaggle competitions in various tasks including medical diagnosis, recommendation systems, and autonomous driving. It has great potential for running on edge devices due to its binary tree-based simple computing kernel, offering unique advantages over deep learning. Despite this, the efficient end-to-end realization is hindered by multiple design challenges due to: 1) the highly irregular tree shape; 2) low hardware utilization; 3) delay from the fully sequential processing of each tree node or high power consumption from the fully parallel processing of all tree nodes; and 4) a large amount of data movement to all the nodes. We propose a low-power and high-performance XGBoost accelerator by employing modular unit trees (MUTs) with reconfigurable interconnects and advanced gating techniques along with selective data movement and execution. The proposed accelerator has been fabricated in a 65-nm process. It achieves 52.5 TOPS/W and 0.41 TOPS/mm², delivering $253{\times }$ and $404{\times }$ higher energy and area efficiencies, respectively, than the state-of-the-art tree-based classifiers.