Hybrid Mixed Signal Computer (HxC)

Project ID: 1617-AP
Available for licensing

Background

Solving differential equations is critical to all science, engineering, and technology. Nature?s behavior and most man-made devices are described by differential equations. Existing numerical methods on digital computers approximate the continuous derivatives of differential equations with discrete differences, which induce unavoidable truncation errors that destabilize the numerical solver. Stability problems are alleviated by shorter time steps, but this increases the computer time required for solution. Stiff and strongly nonlinear systems demand very small time steps. As a result, simulating 1 µs of a CMOS circuit can take weeks, or controllers of systems such as power grids or missiles cannot employ sophisticated models for real time control.

Invention Description

Researchers at University of Texas at Austin have developed intellectual property that allows designing and building a hybrid (mixed-signal: analog & digital) processor for solving systems of ordinary differential equations. The technology merges analog and digital circuitry to overcome computational difficulties of analog and digital computer platforms. The machine consists of a multitude of interconnected, interacting, mixed-signal integrator cells augmented with a mixed-signal nonlinear function approximator. All integrations and calculations are time continuous over the set of real numbers, using a mixed signal floating-point representation with exponent, mantissa and analog bit. The hybrid - not a sampled data system, but true mixed-signal - will substantially outperform modern supercomputers in speed, but at a tiny fraction of the cost, size and power consumption. The new platform permits signal processing without analog-to-digital converters, and could be embodied as a co-processor, giving a laptop or hand held device supercomputer capability. On a benchmark problem of a million coupled differential equations, the HxC solved the equations thousands of times faster than the current fastest supercomputer, making 1 s of the HxC equivalent to 2 hours of that supercomputer.

Benefits

Speed of solution is independent of problem size: the more complex the problem, the larger the improvement over existing technology
Outputs result in traditional floating point or analog, allowing compatibility with existing platforms
Easy integration into any analog, digital, or mixed signal system
Scalable

Features

Computes with real numbers and maintains real number representation - as opposed to digital's traditional floating point, which is actually an integer.
The patented mixed-signal integrator performs time-continuous integrations, maintains a real number representation, and digitizes solutions without A/D converters.
The integrator also avoids analog amp saturation and drift with the aid of digital circuits.
The patented mixed-signal approximator can approximate any arbitrary function in analog hardware, augmented with digital.
To overcome analog precision issues, circuits self-tune integrator constants using the system clock.
Novel layouts of conductors and devices on-chip avoid conductor crossovers and concomitant switching.
To achieve a general solver, the solution applies existing FPGA and FPAA technologies to flexibly interconnect integrator and function approximator blocks, permitting general programmability.
Operating software renders the processor transparent to the end-user.

Market Potential/Applications

Since this invention is compatible with existing technologies, the invention can be applied to any application that requires quickly and accurately solving highly nonlinear, large order differential equations at real-time speeds or faster.

Examples include - predicting power grid responses, automated control systems (such as guidance systems and robotics controls), interactive systems (such as bioelectrical prosthetics interfaces and real time translational systems), diagnostic and prognostic systems (such as predicting critical machine part failures), gaming and animation, defense, aerospace, and climate simulation.

Development Stage

Proof of concept

IP Status

One U.S. patent issued: 7,454,450

UT Researcher

Michael D. Bryant, Ph.D., Mechanical Engineering, The University of Texas at Austin
Benito R. Fernandez-Rodriguez, Ph.D., Mechanical Engineering, The University of Texas at Austin
Ashish R. Seth, Manufacturing System Engineering, The University of Texas at Austin
Shouli Yan, Ph.D., Electrical and Computer Engineering, The University of Texas at Austin
Brian Remy, Mechanical Engineering, The University of Texas at Austin

OTC Contact Information

Jitendra Jain, Licensing Specialist
jjain@otc.utexas.edu
512-471-9055