Efficient Multi-Precision Approximate Posit Multiply-Accumulate Unit
Nuo Si, Qi Wen, Seok‐Bum Ko, Hao Zhang · 2025
In recent years, the posit format has shown significant advantages in machine learning due to its dynamic range adaptability. However, in fields such as scientific computing and signal processing, where both low and high precision operations are required, techniques like multi-precision and mixed-precision are essential to fully harness the potential of posit. Current hardware research primarily focuses on single-precision optimization, leading to challenges in multi-precision scenarios, such as resource wastage and limited flexibility. Moreover, exact multi-precision MAC units still incur high hardware costs. This paper proposes a flexible multi-precision approximate posit MAC unit supporting Posit8, Posit16, and Posit32. By employing the Mitchell approximation algorithm and using a simple piecewise compensation circuit for error correction, the proposed design effectively reduces computational complexity, addresses hardware overhead, enhances computational efficiency, and achieves a balance between performance and resource usage.