Thermodynamic Mixture-of-Experts Gadget for Direct Data Relay

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing machine learning algorithms using classical computing devices face challenges with high execution latency and energy inefficiency due to complex statistical calculations, and thermodynamic computing systems face inefficiencies in communication between devices.

Innovation Solution

Implementing a mixture of experts gadget using thermodynamic chips, comprising a SoftMax energy-based model and an analog summation model, to efficiently process data and relay thermodynamic information directly between components, reducing training costs and maintaining information in a thermodynamic state.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If classical computing devices are used to perform statistical probability calculations, then calculation accuracy can be maintained, but execution time and energy consumption increase significantly

Engineering Contradiction:

Improvecalculation accuracyVSAvoidexecution latency

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The patent replaces classical mechanical computing systems with a thermodynamic computing system that uses physical thermodynamic processes (heat flow, temperature gradients) to perform statistical probability calculations. This substitution of mechanical computation with physical thermodynamic processes enables faster execution while maintaining calculation accuracy through the natural statistical behavior of thermodynamic systems.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent changes the fundamental operating parameters from classical binary logic to thermodynamic states (temperature, heat flow). By representing computational states through thermodynamic parameters and using heat flow to encode probability distributions, the system achieves both speedup and accuracy through the physical laws governing thermodynamic processes.

Inventive Principle:

Principle #35Parameter changes

2Use of energy by moving object

If thermodynamic computing devices are used to perform algorithms, then energy efficiency improves, but communication between devices requires conversion to classical computing form, reducing benefits

Engineering Contradiction:

Improveenergy efficiencyVSAvoidcommunication conversion complexity

Core Design Contradiction:

Use of energy by moving objectVSDevice complexity

Solution Approach 1:

The patent merges multiple thermodynamic computing devices into a unified system where devices can directly communicate through shared thermodynamic fields. Instead of requiring conversion between classical and thermodynamic forms, the system allows direct thermodynamic interaction between devices, eliminating the conversion bottleneck and maintaining energy efficiency advantages.

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The patent introduces thermodynamic fields (heat flow, temperature distributions) as intermediaries that enable direct communication between thermodynamic devices. These thermodynamic intermediaries allow information transfer without requiring conversion to classical computing forms, thus preserving the energy efficiency benefits while enabling multi-device communication.

Inventive Principle:

Principle #24Intermediary (Mediator)

3Use of energy by stationary object

If thermodynamic computing is used for machine learning operations, then training costs and execution energy are reduced, but readout errors and delays may occur

Engineering Contradiction:

Improvetraining energy costVSAvoidreadout accuracy

Core Design Contradiction:

Use of energy by stationary objectVSReliability

Solution Approach 1:

The patent implements feedback mechanisms that continuously monitor thermodynamic states and correct readout errors through iterative refinement. By using feedback from thermodynamic measurements to adjust subsequent computations and readout operations, the system maintains high reliability while preserving the energy efficiency advantages of thermodynamic computing for machine learning training.

Inventive Principle:

Principle #23Feedback

Data Source

PatentUS20250284998A1Mixture of experts energy based model gadget

Publication Date: 2025.09.11 EXTROPIC CORP

AI summary

A thermodynamic mixture of experts gadget includes a SoftMax gadget, multiple energy-based models for processing data, and a summation gadget, also called a Selection of Experts gadget. The SoftMax gadget generates one-hot encoded vectors, which correspond to particular ones of the energy-based models for processing data. The outputs of the energy-based models for processing data, in combination with the one-hot encoded vectors, are inputs to the summation gadget, which generates output that is processed data, processed by energy-based models selected by the SoftMax gadget.