Lattice - Team 11

Problem:

Neural nets are slow and energy hungry.

This is primarily driven by I/O memory costs and the general time complexity of matrix multiplication. This leads to wider problems in a world where Generative AI is increasing in importance. Energy consumption of the AI industry is only increasing, and is driven by not only usage but also more power hungry chips.

Problem Impact

We're building a more time and energy efficient and heat resistant matrix multiplication device. The impact of this solution will be significant on not only current leading Generative AI companies, but also on edge AI applications like agriculture and space. This solution will also be important in the public sector, as many energy grids are unable to sustain the large energy demands from current data centers.

What is it?

We're building a Ferroelectric Diode crossbar array capable of compute-in-memory operations and demonstrate an FP4 neural network capable of recognizing handwritten digits from the MNIST dataset.

Built With

• ferroelectrics