mirror of
https://github.com/luxfi/dex.git
synced 2026-09-24 00:26:31 +00:00
zeekay
c9daf10e22
Delete 3 AI-slop benchmarks from pkg/lx/critical_path_bench_test.go that measured no real work and reported fabricated throughput: - BenchmarkMLXAcceleration: b.Skip + no-op slice index, then fabricated orders/sec - BenchmarkQuantumSignatures (+verifyQuantumSignature): verify == len(sig)==512 && len(pub)==256 - BenchmarkConsensusFinalization (+hashBlock/verifyBlock): byte copy + len(hash)==32 Real benchmarks (CriticalOrderMatching, TradeExecution, RiskCheck, EndToEndLatency, ...) untouched; no imports left unused. Replace fake '434,782,609 orders/sec (GPU/MLX) - 2ns/order - zero-alloc' claims in README.md and LLM.md (traced to archive/_lp108-2026-05-04/mlx_engine.go, a 'pure Go simulation' hardcoding OrdersPerSecond: 1000000.0) with numbers measured first-hand on Apple M1 Max: - C++: 11.88M orders/s (10 threads), 169 ns avg match (p50 125 / p99 292), 5.91M single-thread, 49 ns cancel - Go pkg/lx: 2.2M orders/s, 381 ns/order, 1-2 allocs/op (NOT zero-alloc) - DEX matching stays on CPU; the real GPU win is FHE Metal NTT (23x speedup, N=4096 batch=128).