Efficient management of multidimensional data is critical for high-performance AI tensor and image processing. This paper explores the application of a patented data mapping technique (US7907473, JP5133073), originally designed for semiconductor memory, to improve memory access locality in AI workloads. By efficiently mapping multidimensional tensors according to their coordinates, the approach reduces unnecessary memory activations, potentially lowering power consumption and enhancing computational efficiency. We present a conceptual framework for integrating this method into AI tensor operations, analyze its potential benefits in terms of memory access patterns, and discuss implications for low-power AI system design.
Tatsuya Ishizaki (Tue,) studied this question.