Skip to content

Instantly share code, notes, and snippets.

@docularxu
Last active April 11, 2026 23:57
Show Gist options
  • Select an option

  • Save docularxu/0985c5fbbd32aa3c9fb5e343c1235023 to your computer and use it in GitHub Desktop.

Select an option

Save docularxu/0985c5fbbd32aa3c9fb5e343c1235023 to your computer and use it in GitHub Desktop.

Empowering RISC-V with AI Agents — RVA23 Profile & Practical RISC-V Skills

About

An online meetup hosted by the RISC-V International China community on April 10, 2026. Presented by:

  • Jian Zhang — Open-Source RISC-V CPU Product Manager
  • Guodong Xu — SW Director, RISCstar; RISC-V Ecosystem Practitioner

What We Covered

RVA23 is ready — for the first time, both hardware (SpacemiT K3/V100, T-Head C950, Tenstorrent Atlantis, SiFive P870) and software (openEuler, Ubuntu 26.04, Linux kernel hwprobe) are targeting RVA23 support simultaneously.

The problem: LLMs hallucinate on RISC-V specs. Ask one about the B extension and it'll say "3 instructions" — it's actually Zba + Zbb + Zbs, dozens of instructions across three sub-extensions.

Our solution: Two AI-assisted workflows for daily RISC-V development:

  1. Spec lookup without hallucinations — An Agent Skill built on riscv-unified-db (UDB), a machine-readable YAML database of 2000+ instructions and 400+ extensions. The AI queries structured data instead of guessing.

  2. AI-driven Linux kernel review — Combining three tools to review real kernel patchsets end-to-end:

    • semcode — Tree-sitter semantic code search + lore.kernel.org mail search via MCP
    • review-prompts — Kernel coding conventions encoded as AI skills (/kreview, /kdebug)
    • riscv-spec Skill — Architecture correctness verification against UDB

    Demo scenario: reviewing Qualcomm's RVA23 RFC patchset — searching 34 emails, analyzing 10 call sites via AST, and catching an ISA_EXT vs HWPROBE_EXT discrepancy.

Links

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment