Abstract
A fine mapping strategy is essential for optimizing the layout and execution speed of large-scale neural networks on many-core systems. However, the benefits of many-core systems diminish when applied to neural networks with significant data and computational demands, due to imbalanced resource utilization between space and time when relying on existing single spatial or temporal mapping strategies. To tackle this challenge, we introduce the concept of spatial-temporal density and propose a spatial-temporal density mapping method to fully leverage both spatial and computational resources. Within the framework of the proposed method, we further introduce two approaches: the Negative Sequence Memory Management (NSM) method, which enhances spatial resource (i.e. core memory) utilization, and the Many-core Parallel Synchronous (MPS) approach, which optimizes computational resource (i.e. core multiply and accumulate units, MACs) utilization. To demonstrate the superiority of these methods, the mapping techniques are implemented on our state-of-the-art many-core chip, TianjicX. The results indicate that the NSM method improves spatial utilization by a factor of 3.05 compared to the traditional Positive Sequence Memory Management (PSM) method. Furthermore, the MPS approach increases computational speed by 6.7% relative to the previously widely adopted pipelined method. Overall, the spatial-temporal density mapping method improves system performance by a factor of 1.85 compared to the commonly employed layer-wise mapping method, effectively balancing spatial and temporal resource utilization.