Emulation of Large Language Models for RISC-V using QEMU

• Giovani L. B. Santos UNICAMP
• Lucas Wanner UNICAMP

DOI: https://doi.org/10.5753/eradsp.2025.9700

Abstract

This work aims to execute DeepSeek on a RISC-V emulator, with the objective of showing the viability of executing LLMs in RISC-V devices, as well as in embedded systems. This is important for the purposes of embedding AI, infra independency, energetic efficiency and open source. We used the QEMU emulator to execute RISC-V and obtained the tokens per second count in various tests. Our best result was 1.2415 tks/sec, a value considered slow, albeit viable.

References

Fang, J., Varbanescu, A. L., and Sips, H. (2011). A comprehensive performance comparison of cuda and opencl. International Conference on Parallel Processing, pages 216–225.

Gerganov, G. (2023). llama.cpp. [link].

Moore, S. K. (2023). Risc-v laptops now available. [link].

OpenMathLib (2025). Openblas. [link].

Team, Q. (2025). Qemu documentation. [link].