Mandheling: Mixed-Precision On-Device DNN Training with DSP Offloading

Abstract

This paper proposes Mandheling, the first system that enables highly resource-efficient on-device training by orchestrating the mixed-precision training with on-chip Digital Signal Processing (DSP) offloading. Mandheling fully explores the advantages of DSP in integer-based numerical calculation by four novel techniques: (1) a CPU-DSP co-scheduling scheme to mitigate the overhead from DSP-unfriendly operators; (2) a self-adaptive rescaling algorithm to reduce the overhead of dynamic rescaling in backward propagation; (3) a batch-splitting algorithm to improve the DSP cache efficiency; (4) a DSP-compute subgraph reusing mechanism to eliminate the preparation overhead on DSP. We have fully implemented Mandheling and demonstrate its effectiveness through extensive experiments. The results show that, compared to the state-of-the-art DNN engines from TFLite and MNN, Mandheling reduces the per-batch training time by 5.5$\times$ and the energy consumption by 8.9$\times$ on average. In end-to-end training tasks, Mandheling reduces up to 10.7$\times$ convergence time and 13.1$\times$ energy consumption, with only 1.9%-2.7% accuracy loss compared to the FP32 precision setting.