Step 01
Request test fixtures
Get the manifest with SHA-256 hashes for msmarco-100k, msmarco-1M, and msmarco-10M reference encodings.
Verify · Gen-5 Bit-Identity
The same Gen-5 representation across CUDA, Vulkan, Metal, WebGPU, AVX-512 and NEON — verify the hash yourself. SHA-256 manifests included.
The claim
Top-N produced by the pipeline = Top-N produced by brute-force computation over the corpus. “Equal” means the same set of document IDs, in the same order, with the same distance values. Not high recall. Not approximately equal. Equal — verified per query — across hardware vendors.
Verified fixtures
| Platform | Hardware |
|---|---|
| ARM NEON | Apple M3 Pro |
| x86 AVX-512 | Sapphire Rapids · AMD EPYC 9454P |
| NVIDIA Vulkan | H100 PCIe |
| Apple Metal | M3 Max |
All 9 fixtures byte-for-byte identical. Mechanism: deterministic ChaCha20-based RNG path replacing platform-dependent default hashers (patent-pending).
4-step verification
Step 01
Get the manifest with SHA-256 hashes for msmarco-100k, msmarco-1M, and msmarco-10M reference encodings.
Step 02
aqea-bench encode --corpus msmarco-1M --output corpus.aqea — on your CUDA, Metal, AVX-512, or NEON box.
Step 03
Same command, different vendor. sha256sum corpus.aqea on each side.
Step 04
Hashes should be identical. Reference values for our published runs ship in the manifests/ directory.
Why this is unique
ANN systems (FAISS-HNSW, IVFPQ) cannot publish bit-identical cross-vendor fixtures — they trade recall for speed by design. Exact systems (FAISS-Flat) reproduce only on their native hardware. AQEA is the only system where you can prove byte-equality across hardware vendors at production scale.
SHA-256 manifests are public. If a GitHub repository is not yet linked from this page, request the manifests via the engineering contact below.
Request the test fixtures and reference SHA-256 manifests. We'll send the manifest + the `aqea-bench` access details.