A convertible neural processor supporting adaptive quantization for real-time neural networks

This paper presents Stimpack, a responsive approach toward adaptive neural processing unit (NPU), aiming to satisfy service-level objectives (SLOs) under highly massive loads of neural network (NN) inference. Initially, Stimpack operates in a base-mode processing and computes NNs similarly to a conventional NPU accelerator. During base-mode processing, if SLO violation is likely to occur soon, it changes the operation mode to a burst-mode processing. In the burst-mode processing, Stimpack computes quantized networks instead of the original ones, enabling computational throughput to be scaled up to two times higher than the base-mode processing. This switchable processing is facilitated by three hardware/software schemes. First, a reconfigurable core is adopted to support two different precisions and boost processing throughput for avoiding SLO violations. As computation resources can be shared between the two operation modes, the area overhead of a reconfigurable core is negligible. The second is an on-chip quantization unit that mitigates the data transfer overhead incurred by mode changing. It quantizes parameters stored in on-chip memory on the fly, instead of bringing quantized parameters from off-chip memory. Third, Stimpack leverages a scheduler that determines mode switching based on the amount of workloads in the server. By monitoring ongoing and queued requests of NPU, the scheduler conservatively activates burst-mode processing to minimize accuracy loss. Our analysis shows that, compared to a state-of-the-art NPU, Stimpack achieves 48.4% speedup and allows a 41.4% large load on average while satisfying SLO and near-ideal accuracy.