TernML: A ternary neural network for 50 cents without an FPU
Mikhail T. (Sh0ny)
21 июня 20261 min read
In short
The developer introduced TernML—a neural network with weights limited to {-1, 0, +1} that operates without a floating-point unit. This makes it possible to run AI on low-cost microcontrollers costing about 50 cents.
What is TernML?
TernML is a neural network that uses ternary weights: -1, 0, and +1. This architecture does away with the traditional FPU (floating-point unit), which drastically reduces computational requirements.
Why is this important?
Conventional neural networks require powerful GPUs or specialized accelerators. TernML, on the other hand, can run on simple microcontrollers costing about 50 cents (~36 rubles). This paves the way for embedded AI in IoT devices, sensors, and other low-cost electronics.
From GraphKAN to TernML
The project’s previous version—GraphKAN—achieved 96.15% accuracy on the MNIST dataset while using only 15 KB of memory. TernML took it a step further: the author redesigned the architecture to achieve even more efficient use of memory and computational resources.
Key Features
- No FPU — operations are limited to integers and ternary values.
- Minimal power consumption — suitable for battery-powered devices.
- Compactness — the model fits within tens of kilobytes.
Although specific accuracy figures for TernML are not provided, the overall concept promises interesting applications in edge computing.