Qingming-STARK-G64: a Goldilocks STARK backend on AMD ROCm/HIP

I built and released qingming-stark-g64 , a Goldilocks/G64 STARK backend targeting AMD ROCm/HIP. Repository: https://github.com/uulong950/qingming-stark-g64 The artifact exposes a simple proof pipeline: CLI prover → QSPG64 .qsp proof file → standalone verifier I verified the full retained-Merkle-tree matrix from SCALE20 through SCALE27 . Every scale completed: prover: PASS standalone verifier: PASS The most important practical point is SCALE24 : SCALE24 2^24 rows 1,048,576 trace rows 337 ms proving time 192 MiB retained Merkle tree standalone verifier PASS This looks like a practical application point: million-row STARK proving with sub-second latency on a consumer AMD GPU backend. The flagship benchmark point is SCALE27 : SCALE27 2^27 rows 8,388,608 trace rows 2.05 s proving time 1.5 GiB retained Merkle tree fast_prelayout_xyz standalone verifier PASS The artifact is designed around a concrete backend boundary: source-visible prover explicit proof file independent verifier AMD ROCm/HIP proving path cost-controlled local proving The structural implication is that more applications may be able to move from ordinary backend computation to proof-carrying computation: result → result + proof Business logic, settlement, compliance checks, identity checks, data processing, AI-agent task verification, and decentralized services can expose verifiable outputs instead of only trusted outputs. SCALE27 shows the upper benchmark path. SCALE24 shows the application point. The key message is simple: million-row STARK proofs can become fast enough, small enough, and cheap enough to be called by real applications. 1 post - 1 participant Read full topic