FPnew: An Open-Source Multiformat Floating-Point Unit Architecture for Energy-Proportional Transprecision Computing

The slowdown of Moore's law and the power wall necessitates a shift toward finely tunable precision (a.k.a. transprecision) computing to reduce energy footprint. Hence, we need circuits capable performing floating-point operations on wide range precisions with high proportionality. We present FPnew, highly configurable open-source transprecision unit (TP-FPU), supporting standard custom FP formats. To demonstrate flexibility efficiency FPnew in general-purpose processor architectures, extend RISC-V ISA half-precision, bfloat16, an 8-bit format, as well SIMD vectors multiformat operations. Integrated into 32-bit core, our TP-FPU can speedup execution mixed-precision applications by 1.67× respect FP32 baseline, while maintaining end-to-end reducing system 37%. also integrate 64-bit five formats scalars or 2, 4, 8-way vectors. For this measured silicon manufactured Globalfoundries 22FDX technology across voltage from 0.45 1.2 V. achieves leading-edge efficiencies between 178 Gflop/sW (on FP64) 2.95 Tflop/sW mini-floats), performance 3.2 25.3 Gflop/s.