The Floating Point Trinity: A Multi-modal Approach to Extreme Energy-Efficiency and Performance

The demand for floating-point compute power is ever growing. The domains of big-data, machine learning, and scientific computing require a wide precision range and high operational intensity. The sheer number of operations paired with increased power density implied by technology scaling makes it more important than ever to achieve maximum energy-efficiency for floating point operations. In this work, we present Kosmodrom, our novel silicon solution in Globalfoundries 22 nm Fully-Depleted Silicon on Insulator (FD-SOI) which offers a multi-dimensional approach to trade-off performance, energy-efficiency and power consumption. A variable-precision, dual-core RISC-V system together with a specialized floating point accelerator form the architectural basis. Different implementation strategies and standard cell flavors provide optimal solutions for different operating conditions while supply voltage and forward body bias (FBB) enable for a dynamic trade-off during operation. In this work, we provide a unique insight into the impact of a multitude of tuning parameters to achieve the optimal operating point on the power-performance surface. Kosmodrom achieves a peak energy-efficiency of 260Gflop/s/W and up to 28Gflop/s peak performance within a 6.2-400mW power envelope.