Enabling Extreme High-Throughput Multi-precision Computing on General-Purpose Microprocessor

Deep Neural Networks (DNNs) have shown great acceleration with low-precision computing. To enable low-precision on general-purpose processors for higher computing power, this paper proposes a novel RISC-V core design with multi-precision capability (namely MP-core) but in the flavor of low-precision. We propose two design styles, SIMD and multilevel, to build both integer and floating-point multipliers, and make detailed comparisons regarding their circuit, architecture, and performance beyond DNN applications. With proposed instruction extensions, we show that the SIMD MP-core reserves the single-precision core (namely, SP-core) design with little changes, but only has moderate performance gains. In contrast, the multilevel MP-core flavors low-precision by exploiting the spatial hardware parallelism and temporal instruction-level parallelism to the extreme. With microarchitecture support in the register file and instruction scheduling, the multilevel MP-core improves the performance of linear equation solving by 8.7\(\times \) and 3\(\times \) using different precisions over the SP-core. Our study demonstrates that a general-purpose processor can also have great performance gain from multi-precision computing without loss of generality but flavors low-precision if applications permit.