LLMs on RISC-V

Kernel engineering and performance analysis for LLM inference on RISC-V hardware. This blog documents experiments across both software and hardware modelling — from RVV intrinsics, register pressure analysis, and LLVM IR metrics, to QEMU emulation and cycle-accurate gem5 simulation.

Current focus: Deploying and optimising Liquid AI's LFM's on Milk-V Jupiter.

— Paddy McNabb · AI Engineer · Edinburgh · GitHub · LinkedIn

RISC-V on AI right podcast

2026-04-23 Paddy McNabb

Great discussion with Kris McFadyen and Andy McMahon on AI & RISC-V

[Read more]

Booting Milk-V Jupiter RISC-V Board

2026-04-15 Paddy McNabb

risc-v spacemit milk-v jupiter X60

Flashing BianbuOS Desktop onto the Milk-V Jupiter Board via NVMe SSD

[Read more]

Vectorising convolution kernels in llama.cpp

2026-04-10 Paddy McNabb

risc-v lfm2 llama.cpp rvv hybrid

Experimenting with a range of RISC-V vectorisation ops for comparison to the scalar op in llama.cpp.

[Read more]

Optimising LLM Inference on RISC-V with Custom GEMM Kernels

2026-03-23 Paddy McNabb

risc-v llm gemm llama.cpp rvv gem5 performance

Extending llama.cpp RISC-V GEMM tile from 4×8 to 4×16 — register pressure analysis, LLVM IR metrics, and cycle-accurate gem5 simulation showing a 31% improvement in cycles per output value.

[Read more]

RISC-V LLM Code Reference

2026-03-23 Paddy McNabb

risc-v rvv llama.cpp gem5 qemu code

A concise reference for RISC-V vector intrinsics, QEMU emulation, gem5 simulation, and llama.cpp kernel patterns used in LLM inference optimisation.

[Read more]