Extropic THRML Experiments

Hands-on exploration of Extropic's thermodynamic computing library and probabilistic graphical models.

📚 Documentation

• Extropic_Knowledge_Base.md - Comprehensive overview of Extropic's technology, hardware, and algorithms

🎯 Project Ideas

1. Ising Model Phase Transition Explorer ⭐

Status: ✅ Complete
Difficulty: Beginner-friendly
Directory: ./01_ising_phase_transitions/

Goal: Build an interactive visualization of 2D Ising model phase transitions

Completed Features:

• ✅ Visualize magnetic spin systems on 2D lattices (ising_animated.py)
• ✅ Sweep temperature parameter to observe spontaneous magnetization (phase_transition.py)
• ✅ Animate block Gibbs sampling in real-time (ising_animated.py)
• ✅ Compare different graph colorings and block structures (block_comparison.py)
• ✅ Benchmark sampling speed vs naive methods (benchmark.py)

Scripts:

• ising_basic.py - Simple 1D chain introduction
• ising_interactive.py - Web interface for parameter exploration
• ising_animated.py - Visual 2D grid with spin evolution
• phase_transition.py - Temperature sweep to find critical point
• block_comparison.py - Compare 2-coloring vs 4-coloring strategies
• benchmark.py - Speed comparison: GPU vs CPU vs naive Python

Learning Outcomes:

• ✅ Understand block Gibbs sampling fundamentals
• ✅ See how TSU-style hardware would accelerate sampling
• ✅ Build intuition for energy-based models
• ✅ Classic statistical physics meets modern ML

🔬 Benchmark Results:

We ran comprehensive performance tests comparing THRML (GPU/CPU) against naive Python implementations. Here's what we found:

Grid Size	Naive Python	THRML CPU	THRML GPU	Winner
8×8	3823 s/s	662 s/s	193 s/s	Naive (overhead dominates)
16×16	1003 s/s	685 s/s	178 s/s	Naive (still too small)
32×32	241 s/s	665 s/s	187 s/s	THRML wins! ✨
64×64	60 s/s	564 s/s	218 s/s	THRML (9.4x faster)
128×128	15 s/s	400 s/s	188 s/s	THRML (27x faster)

Key Findings:

• 📍 Crossover point: ~500 spins (between 16×16 and 32×32)
• 🐌 Below this: naive Python wins due to JAX/GPU overhead
• 🚀 Above this: THRML's vectorization dominates (scales much better)
• 🤔 GPU surprisingly slower than CPU for these sizes (memory transfer overhead)
• 💡 Extropic's 10,000x claims target problems with millions of variables, not hundreds

Try it yourself:

cd 01_ising_phase_transitions
python benchmark.py

Challenge: Can you find ways to make THRML faster? Ideas:

• Better JIT warmup strategies
• Optimize batch sizes
• Test larger grids (256×256, 512×512)
• Try different blocking strategies
• Profile memory vs compute bottlenecks

---

2. Energy-Based Generative Model on MNIST 🎨

Status: 📋 Planned
Difficulty: Intermediate
Directory: ./02_ebm_mnist_generation/

Goal: Train and sample from an EBM to generate handwritten digits

Features:

• Train EBM on MNIST or Fashion-MNIST dataset
• Sample new images using THRML
• Visualize energy landscape and sampling trajectories
• Benchmark against VAE/diffusion baselines
• Attempt to reproduce Extropic's 10,000x efficiency claims

Learning Outcomes:

• Hands-on with energy-based generative models
• Test claims about sampling efficiency
• Compare EBMs vs current SOTA architectures
• Understand training challenges for EBMs

---

3. Probabilistic Constraint Satisfaction Solver 🧩

Status: 💡 Future
Difficulty: Advanced
Directory: ./03_constraint_satisfaction/

Goal: Map combinatorial optimization problems to Ising models and solve with Gibbs sampling

Features:

• Implement Sudoku, graph coloring, or scheduling problems
• Map constraints to Ising/Potts model energy functions
• Use THRML sampling to find solutions
• Compare vs traditional SAT/CSP solvers
• Visualize solution space exploration

Learning Outcomes:

• Non-ML applications of probabilistic computing
• When is Gibbs sampling competitive with other methods?
• Problem formulation for thermodynamic hardware
• Understand TSU applicability beyond AI

---

🛠️ Setup

# Create virtual environment
python3 -m venv .venv
source .venv/bin/activate

# Install THRML and dependencies
pip install thrml jax jaxlib matplotlib numpy

# For CUDA support (optional, if you have NVIDIA GPU)
pip install --upgrade "jax[cuda12]"

📖 Resources

• THRML GitHub: https://github.com/extropic-ai/thrml
• Documentation: https://docs.thrml.ai
• Extropic Website: https://extropic.ai
• Paper (DTMs): https://extropic.ai/dtms

🧪 Experimental Philosophy

This workspace is about hands-on validation of Extropic's claims:

• Does block Gibbs sampling actually provide speedups?
• Are EBMs competitive with modern architectures?
• Where does thermodynamic computing shine vs struggle?
• Can we reproduce their published results?

Approach with curiosity and healthy skepticism! 🔬