Trans-precision NPU for resource-efficient mobile AI acceleration

Abstract

This paper introduces the Trans-precision NPU (T-NPU), a resource-efficient, multi-mode neural processing unit for mobile AI acceleration using a unified multi-precision datapath. At its core is a Single Instruction Multiple Data (SIMD) trans-precision multiply-accumulate unit (T-MAC) supporting heterogeneous formats, including FP4/FP8, INT4/INT8, and Posit-8/16, enabling flexible computation with improved throughput. The design uses a reconfigurable datapath that shares logic between exponent and mantissa processing to boost arithmetic intensity and energy efficiency. T-MAC integrates the DA-VINCI activation function to form a fundamental processing element for diverse workloads such as DNNs, Transformers, reinforcement learning, and generative AI. We evaluate an AXI-enabled unified core combining a systolic array and vector engine, integrated with a RISC-V CVA6 core and memory via the Cheshire framework. Results show a 14 pJ PDP, 2.4× better energy efficiency in 28 nm CMOS, and 1.8× fewer FPGA LUTs. MobileNet-SSD with KALAM-16 encoding shows ∼0.6% accuracy loss and 4.47 TOPS/W. © 2026 Elsevier B.V.