Back close

SPM-Arrow: A Memory-Hierarchy Optimized RISC-V Vector Accelerator for Energy-Efficient Machine Learning

Publication Type : Conference Paper

Publisher : IEEE

Source : 2026 IEEE International Conference on Interdisciplinary Approaches in Technology and Management for Social Innovation (IATMSI)

Url : https://doi.org/10.1109/iatmsi68868.2026.11466272

Campus : Amritapuri

School : School of Engineering

Center : Humanitarian Technology (HuT) Labs

Department : Electronics and Communication

Year : 2026

Abstract : When making an ML inference about micro- or edge devices, the need to either spend many units of compute, and, to ensure that the battery charge is low. The vectors such as RISC -v architecture such as Arrow are rapidly unaccelerating matrix -vector. Once again consume enormous amounts of off -chip DDR data, which kills power efficiency and introduces immense bottlenecks in convolution and attention. Suchly we came up with SPM-Arrow: RISC-V based scalar/vector accelerator, which includes a scratchpad (which appears to be dual-bank), hardware controlled double buffering and a DMA controlled tiling engine, which is executed in a deterministic manner. It also works on this cache-centricized architecture but unlike the traditional designs, SPM-Arrow offers the predictable low-latency access and merely the optimum possible data reuse as overlapping moves of compute and asynchronous memory, all RISC-V Vector Extension (RVV)-compatible. our working prototype was built around Xilinx Artix-7 board which has FPGA and values are very exciting. The SPM-Arrow occupies up to 3.2× the bandwidth of the DDR, is some 2x faster and consumes some 68% less power than an otherwise standard Arrow like accelerator. It would imply 2.8 times in terms of performance/watt and the extra hardware would add on to the LUT usage at another 10 per cent. The simulation waveforms testify to the fact that a double buffering is ideal and the scratchpad swaps is being executed safely and we are right around 94 per cent of the vector units in use. All in all, these results provide insight into why close compute-memory coupling is an issue that should be considered with regard to next-generation RISC-V ML accelerators. Unscaling Simple predictable memory schedules like SPM-Arrow provide a realistic solid scalable architecture which can deliver the real-time performance required on (traditional) nets of edge inference applications, either convolutional nets or transformer kernels or signal-processing pipelines.

Cite this Research Publication : Kisor Shanmuga Sundaram, Arun Kirthick Chithambaram, Rajesh Kannan Megalingam, SPM-Arrow: A Memory-Hierarchy Optimized RISC-V Vector Accelerator for Energy-Efficient Machine Learning, 2026 IEEE International Conference on Interdisciplinary Approaches in Technology and Management for Social Innovation (IATMSI), IEEE, 2026, https://doi.org/10.1109/iatmsi68868.2026.11466272

Admissions Apply Now