Ternary Neural Processing Unit

Overview

A neural processing unit built from arrays of ternary multiply-accumulate cells, each incorporating a photonic-ternary computing device with a ferroelectric weight shell. Neural network weights drawn from the ternary set {+1, 0, −1} are stored as polarisation states of that shell and programmed optically — there is no electrical weight-loading interconnect.

The central result is that multiplication by a ternary weight requires no multiplier circuit at all. A weight of +1 routes the input trit through unchanged. A weight of −1 inverts it, which in balanced ternary is a wire crossing. A weight of 0 withholds the optical pulse entirely, leaving the device dark and consuming no switching energy.

That last case is the important one. In a binary accelerator a zero weight still costs a multiply and still burns energy; sparsity has to be utilized by software. Here a zero weight is physically an absence of light, so sparsity is structural — it saves energy whether or not anything schedules around it.

Key Innovations

• Multiply-accumulate without a multiplier circuit
• Ternary weights stored as ferroelectric polarisation states, programmed optically
• No electrical weight-loading interconnect
• Weight zero is a dark device — structural sparsity at zero switching energy
• Accumulation by current summation rather than an adder tree
• Native match between ternary hardware and ternary-quantised neural networks

Why It Matters

Ternary weight quantisation is already where a great deal of machine-learning inference is heading, because networks quantised to {+1, 0, −1} retain most of their accuracy at a fraction of the memory cost. Those networks currently run on binary hardware that must emulate ternary arithmetic.

This patent inverts that: the hardware is natively ternary, so a ternary-quantised network maps onto it directly. Cell counts, energy figures, and array dimensions are shared under mutual NDA.

Disclosure

This page describes the architecture and purpose of the invention. Specific parameters — dimensions, thresholds, wavelengths, code assignments, and simulation figures — together with the complete claim set are shared under mutual NDA. Contact manish@manitlab.org to request access.

Related Patents