benchmarks/README.md

Sofia Engine Empirical Performance Benchmarks

All benchmark metrics published below represent real measurements executed on physical test hardware using benchmarks/run_benchmarks.py. No metrics are fabricated or estimated.

---

1. Hardware Profile A Reference Measurements

• OS / Platform: Windows 11 (AMD64)
• CPU: Intel64 Family 6 Model 165 (x86_64)
• Python Version: 3.12.10 CPython
• Timestamp: 2026-09-21 23:56:41 UTC

Latency Percentiles ($N = 100$ iterations)

Benchmark Task	Target Budget	Mean	$p_{50}$ (Median)	$p_{95}$	$p_{99}$	Budget Status
Package Import Time	$\le 500\text{ ms}$	$342.8\text{ ms}$	—	—	—	PASS
FFT / Welch PSD (1024 samples)	$\le 500,\mu\text{s}$	$319.2,\mu\text{s}$	$307.6,\mu\text{s}$	$362.8,\mu\text{s}$	$445.5,\mu\text{s}$	PASS
FFT / Welch PSD (4096 samples)	$\le 2000,\mu\text{s}$	$430.3,\mu\text{s}$	$387.4,\mu\text{s}$	$597.3,\mu\text{s}$	$713.1,\mu\text{s}$	PASS
Analytic Envelope (2048 samples)	$\le 1000,\mu\text{s}$	$141.0,\mu\text{s}$	$127.8,\mu\text{s}$	$200.2,\mu\text{s}$	$237.9,\mu\text{s}$	PASS
Symmetrical Components (Fortescue)	$\le 1500,\mu\text{s}$	$60.7,\mu\text{s}$	$55.5,\mu\text{s}$	$93.0,\mu\text{s}$	$102.7,\mu\text{s}$	PASS
14-Feature Extraction (1024 samples)	$\le 10000,\mu\text{s}$	$2521.5,\mu\text{s}$	$2415.5,\mu\text{s}$	$3304.8,\mu\text{s}$	$3491.0,\mu\text{s}$	PASS
Sofia VM Forward Inference (4→8→1)	$\le 5000,\mu\text{s}$	$1751.2,\mu\text{s}$	$1675.8,\mu\text{s}$	$2210.2,\mu\text{s}$	$2457.1,\mu\text{s}$	PASS
Sofia VM Training Step (analytical backprop)	$\le 20000,\mu\text{s}$	$6609.7,\mu\text{s}$	$6442.9,\mu\text{s}$	$7483.0,\mu\text{s}$	$9051.6,\mu\text{s}$	PASS

Memory Consumption

• Peak traced memory during benchmark suite: $7.14\text{ MB}$ ($\le 32\text{ MB}$ budget for Profile A gateways).