metadata
license: mit
tags:
- pytorch
- safetensors
- threshold-logic
- neuromorphic
- arithmetic
- dsp
threshold-saturating-adder
4-bit unsigned saturating adder. Clamps to maximum value on overflow instead of wrapping around.
Circuit
A[3:0] ──┐
├──► Sat Add ──┬──► S[3:0] (result)
B[3:0] ──┘ └──► saturated (overflow flag)
Behavior
| A + B | Result | Saturated |
|---|---|---|
| ≤ 15 | A + B | 0 |
| > 15 | 15 | 1 |
Examples
7 + 5 = 12, saturated=0
10 + 10 = 15, saturated=1 (true sum: 20)
15 + 15 = 15, saturated=1 (true sum: 30)
8 + 7 = 15, saturated=0
8 + 8 = 15, saturated=1 (true sum: 16)
Wrapping vs Saturating
Wrapping: 15 + 1 = 0 (modulo 16)
Saturating: 15 + 1 = 15 (clamped)
Architecture
| Component | Count | Neurons |
|---|---|---|
| Full Adders | 4 | 28 |
| Saturation MUXes | 4 | 12 |
Total: 40 neurons, 124 parameters, 4 layers
How It Works
1. Compute A + B with standard 4-bit adder
2. If carry out = 1:
- Output 1111 (15)
- Set saturated = 1
3. Else:
- Output sum bits
- Set saturated = 0
Applications
- Audio DSP: Prevents clipping distortion
- Image processing: Pixel value saturation
- Neural networks: Activation clamping
- Control systems: Prevents actuator overshoot
- Fixed-point arithmetic: Overflow protection
Usage
from safetensors.torch import load_file
w = load_file('model.safetensors')
# All 256 test cases verified
# Saturates 120 cases (where A+B > 15)
Files
threshold-saturating-adder/
├── model.safetensors
├── create_safetensors.py
├── config.json
└── README.md
License
MIT