Madhava V17 — BIGANN-100M Streaming Stress Test & Benchmark (Verified, int8, Zero Bound Violations)

Madhava V17 — BIGANN-100M Streaming Benchmark (Stress Test, Verified)

C++ implementation with Modified Gram-Schmidt orthogonal projections, int8 quantization with dynamic dequantization margin, and empirical bound violation verification against exact cosine similarity. Zero bound violations across all scales (100K to 100M).

Results (28-thread CPU, AVX2+FMA, 100M vectors)

• Build: 21.5s
• Query latency: 1.9s
• Memory: 18.6GB (int8 projection cache pr1_i8+pr2_i8+e1/e2)
• NDCG@10: 1.000, R@10: 1.000
• Bound violations: 0% (verified empirically)
• Early exit: 124 cosine computations per query (from 2000 candidates)

Kaggle Notebook (live execution)

https://www.kaggle.com/code/kleniopadilha/madhava-v12-bigann-verified

Files

• bigann_stream.cpp — C++ V12 source (MGS + int8 + bound verification + early exit)
• madhava_v12_verified.ipynb — Kaggle notebook with dynamic CSV parsing & Matplotlib plots
• bigann_bench.cpp — Reference benchmark implementation

BSL 1.1 | pay@winnex.ai