ReNPU: A Resource-efficient Multi-Mode Neural Processing Unit with Unified Multi-Precision Datapath for Mobile AI Workloads

This paper discusses ReNPU for mobile AI workloads, a resource-efficient, multi-mode neural processing unit utilizing an unified multi-precision datapath. The proposed Re-NPU consists of a novel SIMD Re-MAC which supports FP4/FP8, INT4/INT8, and Posit-8/16, thus enabling flexible and SIMD computations with enhanced throughput. Reconfigurable datapath between exponent and mantissa processing contributes to increased arithmetic intensity and energy efficiency. The Re-MAC has been integrated with DA-VINCI AF, forming the ReNPU for fundamental Unified core, capable of accelerating mobile AI workloads like DNNs, Transformers, RL, and Gen-AI models. Further, Empirical evaluation on AXI-enabled unified core consisting both systolic array and vector engine, integrated with RISC-V CVA6 core and memory via Cheshire framework. ReNPU demonstrates 14 pJ Power-Delay Product (PDP), 2.4× improvement in energy efficiency at 28nm CMOS, and 1.8× reduction in LUTs for FPGA implementation, compared to stateof-the-art MAC compute engines. The test-case for MobileNet-SSD with KALAM-16 encoding shows just 0.6% accuracy drop, and 4.47 TOPS/W energy efficiency, confirming ReNPU's applicability in next-generation mobile AI SoCs.