2026-10-11 18:04 UTC

Independent testing will determine whether the released open RK3588 NPU compiler and runtime can reliably run GPT-2, SigLIP, and models exported from PyTorch, ONNX, or JAX without Rockchip’s vendor SDK at practically useful performance.

state: expiredheat: lowuncertainty: highconvergesscott: mediumedge-inference open-npu-compilers local-inferenceone_does_not_justRockchip

What is this?

An independent engineer reports reverse-engineering the Rockchip RK3588 NPU and building an open compiler/runtime intended to bypass Rockchip’s SDK, including support claims for GPT-2 and SigLIP. The supplied snippets establish that the RK3588 can accelerate converted models and that PyTorch models commonly follow a PyTorch → ONNX → RKNN workflow, while one author reports running SigLIP through manual hardware-level work. They do not independently verify the released open stack, its claimed 36 tok/s GPT-2 performance, broad PyTorch/ONNX/JAX compatibility, or operation entirely without vendor components.

Why it matters to Scott

The open compiler/runtime converges with Scott’s Sovereign Software Assurance position by attempting to make accelerator operation independent of the vendor SDK, while directly extending his hardware-aware local-inference work to a new edge NPU. If its compatibility and performance claims survive independent testing it could expand practical deployment choices; until then, the evidence is too preliminary for high relevance.
ip:framework.sovereign-software-assurancedev:concept.hardware-aware-local-inferenceip:concept.vendor-lock-inradar:apple-neural-engine-training-accessradar:concept.edge-inferenceradar:concept.local-inferenceradar:concept.ai-hardware
queries asked of Scott's wikis
  • open compiler stacks for proprietary AI accelerators
  • local inference without vendor SDKs
  • edge NPU portability and model conversion
  • hardware sovereignty through reverse engineering
  • local inference performance versus deployment friction
  • open tooling for heterogeneous AI hardware

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (2) — ⭐ canonical anchor

sourceobjectauthorscorecomments
🟠 redditReverse-Engineering the RK3588 NPU: Building an Open Compiler to Run GPT-2 at 36 tok/s
LocalLLaMA
one_does_not_just3521
🟧 echo.blog ⭐The author reports reverse-engineering the RK3588 NPU register format and releasing an open compiler and runtime that run GPT-2 and SigLIP fone_does_not_just——

Interpretation history

Decision trace