Qingming Foundation is the central foundation repository for completed Qingming-series primitive artifacts.
The Qingming series establishes both a common language for measurement and compelling performance evidence.
It collects only artifacts that have reached a Qingming-style level: explicit contract, independent validation, benchmark boundary, implementation, and repository-level reproducibility.
Qingming is not a leaderboard and not a collection of demos. It is a foundation layer for turning low-level primitive performance claims into clearly defined, validated, and comparable new baselines.
Qingming Foundation only lists completed Qingming-style primitive artifacts.
This repository does not list demos, experiments, roadmaps, internal prototypes, or ordinary implementations. A project may enter Qingming Foundation only after it reaches the minimum Qingming-style bar: it must define the structure of a primitive benchmark, provide a validated implementation, report a clear performance boundary, and establish a new baseline that can be used as a foundation for future work.
A listed artifact must provide:
- a clear primitive definition
- an explicit benchmark structure
- an explicit benchmark contract
- an implementation
- correctness or validation checks
- benchmark results
- hardware environment
- build / run path
- clear supported and unsupported boundaries
- repository-level reproducibility
The minimum bar for inclusion is not implementation alone.
A Qingming Foundation entry should be a foundation-moving artifact in its own primitive domain: it should clarify how the primitive is structured, how it should be measured, what boundary the implementation reaches, and how others can reproduce or challenge the result.
Unreleased, experimental, internal, not-yet-restructured, or non-Qingming-style projects are not listed here.
No contract, no comparison.
No correctness, no benchmark.
No validation, no foundation entry.
A Qingming-style artifact should make the following explicit:
- mathematical object
- input / output contract
- layout contract
- timing region
- correctness checks
- hardware environment
- benchmark boundary
- supported / unsupported scale
- reproduction path
| Artifact | Domain | Contract | Implementation | Benchmark | Validation | Status |
|---|---|---|---|---|---|---|
| Qingming-G64-NTT-CUDA | Native Goldilocks/G64 STARK-LDE NTT | Yes | NVIDIA CUDA | Yes | Yes | Public |
| Qingming-G64-NTT | Native Goldilocks/G64 STARK-LDE NTT | Yes | AMD HIP / ROCm | Yes | Yes | Public |
| Qingming-STARK-G64 | Goldilocks/G64 STARK proving backend | Yes | AMD HIP / ROCm | Yes | Yes | Public |
Validation means that the artifact has passed author-side independent validation, such as correctness checks, CPU reference comparison, roundtrip tests, fixed benchmark runs, or release-suite checks.
External third-party reproduction is a separate future status and is not required for initial inclusion in Qingming Foundation.
G64-NTT is the first completed Qingming primitive line.
It focuses on native Goldilocks/G64 STARK-LDE NTT and defines benchmark contracts around field semantics, domain size, layout, timing region, correctness, hardware boundary, and reproducibility.
Domain. Native Goldilocks/G64 STARK-LDE NTT on NVIDIA CUDA.
What this work does. This artifact defines two CUDA benchmark contracts:
qingming_fast: device-native tiled input to mapped outputqingming_standard: natural-order input to standard materialized output
The fast path treats device layout as part of the algebraic operator and exposes mapped output as the proof-facing result. The standard path provides natural-order compatibility and includes layout transition cost.
Boundary.
RTX4090-24G covers domain_logn=20..30.
Representative reference points:
domain_logn=24 fast ~1.26 ms
domain_logn=27 fast ~9.18 ms
domain_logn=30 fast ~97 ms
domain_logn=30 standard ~325 ms
domain_logn=31 is capacity-gated on RTX4090-24G.
domain_logn>=32 is unsupported.
Repository. https://github.com/uulong950/qingming-g64-ntt-cuda
Domain. Native Goldilocks/G64 STARK-LDE NTT on AMD HIP / ROCm.
What this work does. This artifact provides an AMD-native G64 NTT implementation with fast mapped output, standard compatibility path, correctness checks, CPU reference validation, and large-domain scaling.
Boundary.
RX 7900 XTX reaches domain_logn=27 with reported fast-path effective logical bandwidth around 559 GB/s under the Qingming G64/STARK-LDE benchmark contract.
Repository. https://github.com/uulong950/qingming-g64-ntt
G64-STARK is the next Qingming foundation line above the G64-NTT primitive layer.
It focuses on a reproducible Goldilocks/G64 STARK proving and verification backend, with an explicit proof-file boundary, deterministic protocol contract, standalone verifier, and repository-level reproducibility.
Domain. Goldilocks/G64 STARK backend on AMD HIP / ROCm.
What this work does. This artifact provides a source-visible, CLI-based QSPG64 STARK proving and verification backend. It defines a small official integration surface:
- CLI prover
- QSPG64
.qspproof file - standalone verifier
The backend binds public inputs, statement digest, trace commitment, quotient commitment, Merkle openings, FRI proof material, and local AIR verifier material into a complete QSPG64 proof boundary.
Contract. The artifact fixes the Goldilocks/G64 field contract, Poseidon2-G64 hash contract, compiled AIR profile, proof format, retained-Merkle-tree proof-builder behavior, and standalone verification checks.
Boundary. RX 7900 XTX 24GB is the verified backend target. The retained-tree proving matrix is verified from SCALE20 through SCALE27, with SCALE27 as the primary FAST XYZ benchmark.
Repository. https://github.com/uulong950/qingming-stark-g64
Qingming Foundation is the primitive foundation layer.
Qingming = primitive contract + benchmark + boundary implementation
Application-layer projects may use Qingming primitives to build products, PoCs, or workflows.
Application layer = product / PoC / workflow built on Qingming capabilities
Application-layer results should not replace primitive-level benchmark contracts.
Qingming Foundation aims to organize completed Qingming-series artifacts that do three things at once:
- define the benchmark structure of a primitive
- implement the primitive near the hardware boundary
- provide an independently validated baseline for future work
The goal is not to claim isolated performance numbers.
The goal is to establish new ground truth.