FPU Reduced Variable Precision in Time: Application to the Jacobi Iterative Method

International audienceFloating-Point (FP) computation using standard IEEE formats has a significant hardware overhead. Moreover, these formats are over-designed for most real-world applications, especially iterative refinement algorithms. Hence there is a need for hardware FP Unit (FPU) architectures with run-time variable precision capabilities.In this work, we propose a FPU architecture that enables designers to dynamically tune FP computations’ precision at run-time, leading to significant power consumption, execution time, and energy savings. Despite its additional circuit area overhead, the proposed architecture simplifies the integration of variable precision in existing software workloads at any level of the software stack (OS, RTOS, or application-level) since it only requires light-weight software support and solely relies on traditional assembly instructions, without the need for a specialized compiler or custom instructions.We propose a modified version of the Jacobi iterative method with automatic run-time variable precision, which takes full advantage of the suggested FPU. It demonstrates up to 75.21% power consumption saving, up to 64.37% execution time saving, and up to 88.96% total energy saving w.r.t the reference double-precision implementation, and with no accuracy loss.