Synthesis Time Reconfigurable Floating Point Unit for Transprecision Computing

This paper presents the design and the implementation of a fully combinatorial floating point unit (FPU). The FPU can be reconfigured at implementation time in order to use an arbitrary number of bits for the mantissa and exponent, and it can be synthesized in order to support all IEEE-754 compliant FP formats but also non-standard FP formats, exploring the trade-off between precision (mantissa field), dynamic range (exponent field) and physical resources. This work is inspired by the consideration that, in modern low power embedded systems, the execution of floating point operations represents a significant contribution to energy consumption (up to 50% of the energy consumed by the CPU). In this scenario, the adoption of multiple FP formats, with a tunable number of bits for the mantissa and the exponent fields, is very interesting for reducing energy consumption and, simplifying the circuit, area and propagation delay. Adopting multiple FP formats on the same platform complies with the concept of transprecision computing, since it allows fine-grained control of approximation while meeting the required constraints on the precision of output results. The designed FPU has been tested in order to evaluate the correctness of all supported operations, and implemented on a Kintex-7 FPGA. Experimental results are provided, illustrating the impact and the benefits derived by the use of non-standard precision formats at circuit level.