13 November 2018 in IEEE Transactions on Circuits and Systems II: Express Briefs
19 August 2025 in IEEE Access
In today’s technological landscape, embedded and IoT devices face escalating demands for performance and power efficiency in inference tasks employing Convolution Neural Networks (CNNs). Various methods have emerged to improve performance, including quantization and the adoption of fixed-point formats instead of floating-point formats. RISC-V Instruction Set Architecture (ISA) being open-source and with a modular specification poses as a suitable platform for targeting such applications for embedded processing. The RISC-V’s standard P-extension represents a significant advancement in integer-based computation, offering new ways for achieving higher speedups through integer-packed SIMD instructions. This work presents the first RISC-V processor based on RV32IM base ISA with an optional support of P-extension Version 0.9.11-draft-20211209 called the P-CORE. The P-CORE implements P-extension ISA for RV32 specification. The implemented processor is evaluated for Xilinx Zynq Ultrascale+ FPGA using different test case functions; matrix multiplication, convolution, max-pooling, and fully connected. A CNN used for detecting handwritten digits called Lenet-5 is also used as a test case to evaluate the P-CORE. The evaluation also presents Design Space Exploration (DSE) for P-extension by exploiting the level of data-level-parallelism; SIMD8 (8-bit (4-way)) and SIMD16 (16-bit (2-way)). Our investigation demonstrates that with the P-extension activated, CNN algorithm computations can achieve speedups of up to 17x for max-pooling, 7x for matrix multiplication, and 4.8x for fully connected networks, accompanied by similar enhancements in power efficiency. The processor achieved a maximum performance of 335.1 MIOPS and 490.6 MIOPS/W efficiency for matrix multiplication using -O3 compiler optimization.
1 July 2025 in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
Since dynamic power has a quadratic relationship with voltage, reducing voltage is an effective way to lower power consumption in digital circuits. However, maintaining stable operation at lower voltages is challenging due to increased sensitivity to process, voltage, and temperature (PVT) variations, making it difficult to determine optimal operating points using static timing analysis (STA). While circuit- and device-level solutions like timing error detection (TED) systems can enable lower voltage operation, they introduce significant overhead and design complexity. In this article, we integrate an algorithm-based fault tolerance (ABFT) method into the structure of systolic arrays (SAs) to capture timing errors when voltage is scaled down, ensuring safe and optimized low-voltage operation. Our proposed approach, SCISSORS, demonstrates how extra voltage margins in SAs used for matrix arithmetic can be trimmed by integrating a simple algorithmic technique into the structure of the array. This solution not only detects errors in the accelerator but also those caused by voltage reduction in on-chip memory and auxiliary circuits. It is fully implementable through hardware description languages (HDL) without requiring transistor- or circuit-level modifications to the netlist. Implementation on a Zynq system-on-chip (SoC) shows that SCISSORS introduces only a tolerable overhead of 11% and 8% for $32{\times }32$ and $64{\times }64$ SAs, respectively, while achieving nearly a $2{\times }$ improvement in energy efficiency. Experimental results further demonstrate that SCISSORS adaptively adjusts voltage in response to the voltage-temperature coupling behavior of digital circuits at runtime, specifically addressing inverse temperature dependence (ITD).