AXON-RT v0.1.2

Adaptive eXecution Optimization Network - Runtime

A dtype-aware numerical optimization runtime that delivers free performance gains by understanding cache physics.

🎯 The Discovery

We proved that optimal matrix shapes depend on dtype:

dtype	Size ~2048	Speedup	Why
float32	2048 (stay)	—	128 cache lines (aligned)
float64	2056	+9.1%	257 cache lines (breaks conflict)

This is the first runtime to encode this physics.

📊 Validation Results

Test	Result
Unit tests	8/8 passed
Quantum Dive	dtype hypothesis confirmed
Endurance (266 laps)	94.7% win rate, +9.1% avg
BLAS	OpenBLAS 0.3.29 (Haswell kernel)

🚀 Quick Start

from axon_rt import AxonRT
import numpy as np

# Initialize (loads profile, ~10ms)
rt = AxonRT.auto_detect()

# Get dtype-aware shape recommendation (~0.01ms)
rec = rt.recommend_shape((2048, 2048), dtype=np.float64)
print(rec.recommended)  # (2056, 2056)

rec = rt.recommend_shape((2048, 2048), dtype=np.float32)
print(rec.recommended)  # (2048, 2048)

# Get dtype recommendation for workload
dtype_rec = rt.recommend_dtype("reflex")  # np.float32
dtype_rec = rt.recommend_dtype("accumulation")  # np.float64

📁 Package Contents

axon_rt/
├── core/
│   ├── axon_rt.py          # Main runtime (v0.1.2)
│   └── __init__.py
├── tests/
│   ├── test_axon_rt.py     # Unit tests
│   ├── quantum_dive.py     # dtype hypothesis test
│   └── endurance_test.py   # Stability test
├── benchmarks/
│   ├── benchmark_hardware.py
│   └── comprehensive_test.py
├── profiles/
│   └── alienware_profile.json
├── README.md
├── LICENSE
└── requirements.txt

🔧 Installation

# Clone or extract the package
cd axon_rt

# Install dependencies
pip install -r requirements.txt

# Run tests
python -m pytest tests/

# Or run individual test
python tests/test_axon_rt.py

💻 CLI Usage

# Show status
python -m axon_rt status

# Get shape recommendation
python -m axon_rt recommend-shape 2048 2048 --dtype float64
# Output: (2056, 2056)

# Get dtype recommendation
python -m axon_rt recommend-dtype reflex
# Output: float32

# Generate hardware profile
python -m axon_rt profile --out my_hardware.json

🔬 The Physics

Cache Line = 64 bytes

float32 (4 bytes/element):
  2048 × 4 = 8192 bytes/row = 128 cache lines ✓ (aligned)

float64 (8 bytes/element):
  2048 × 8 = 16384 bytes/row = 256 cache lines ✗ (conflict!)
  2056 × 8 = 16448 bytes/row = 257 cache lines ✓ (broken)

When row size is an exact multiple of cache associativity, different rows compete for the same cache lines. Adding 8 elements (2056) breaks this pattern.

📈 Performance

Validated on Alienware m15 R7 (AMD Ryzen 9 6900HX, OpenBLAS):

• float64 @ 2056: +9.1% faster than 2048 (sustained over 266 tests)
• float32 @ 2048: No change needed (already optimal)
• float32 speedup: 1.92x faster than float64

📄 License

MIT License - See LICENSE file

---

Created by: AXONLang Labs CEO Daniel N Santiago AXONLangLABS.com .AXON PROGRAM WILL BE REVEALED SOON! (LOOKING FOR A GOOD HOME FOR IT SO I CAN JUST HAVE MORE TIME TO CREATE)

"Measure YOUR hardware. Apply physics-based rules. Get free speedups."