Timing-based computer architecture systems and methods
Inventors
Madhavan, Advait • Daniels, Matthew William • Stiles, Mark David
Assignees
University of Maryland College Park • United States Department of Commerce
Publication Number
US-11817173-B2
Publication Date
2023-11-14
Expiration Date
2041-10-01
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Abstract
Systems and methods are provided herein that implement and utilize temporal state machines and other temporal computational/memory circuits. For example, a temporal state machine is provided that comprises a temporal computational unit comprising at least one circuit to perform a time-domain operation, and a temporal memory unit. Both the temporal computational unit and temporal memory unit can operate using solely, or partially, time-encoded wavefronts. The computation unit accesses time-encoded wavefronts from the memory and can output and write time-encoded wavefronts back to the memory.
Core Innovation
The invention provides systems and methods to implement computational and memory components using temporal encoding, specifically employing temporal state machines and temporal memory units that utilize time-encoded wavefronts. The temporal computational unit performs time-domain operations accessing and writing time-encoded wavefronts to and from a temporal memory unit, enabling computation and storage directly in the time domain rather than traditional digital clocked schemes.
The problem addressed arises from the stagnation in improvements of classical digital, clocked computational components, compounded by the slowing progress in transistor design and performance. Current designs encode multi-bit data using multiple digital transitions, which consume significant energy. The invention introduces and systematizes race logic encoding, where information is encoded in the timing (arrival time) of single digital edges, reducing energy consumption and simplifying certain computational operations.
Prior to this invention, race logic had been applied in an ad hoc fashion and lacked practical architectures incorporating memory and compositional frameworks to build higher-order temporal computational functions. This invention fills this gap by providing practical, systematized architectures for temporal memories that can be programmed and recalled in the time domain and temporal state machines capable of combining memory and computation using temporal encoding. These enable lifting the causality and invariance restrictions of pure race logic, allowing construction of complex and general-purpose temporal computing architectures.
Claims Coverage
The claims define inventive features involving a temporal state machine integrating temporal computational units and temporal memory units electrically coupled to perform time-domain operations with time-encoded wavefronts.
Temporal state machine integrating time-domain computation and memory
A temporal state machine comprising a temporal computational unit configured to perform time-domain operations, a temporal memory unit, and circuitry to electrically couple the memory input and output lines with the computational unit, enabling access to time-encoded memory recall wavefronts and output of time-encoded wavefronts.
General computational and algebraic functions in temporal domain
The temporal computational unit is configured to perform general computational functions, including min and max functions, as well as algebraic operations, within the time domain.
Electrical connections enabling read and write operations between computational and memory units
Electrical connections are arranged so that computational output lines connect to memory input lines allowing writing temporal data to memory, and computational input lines connect to memory output lines allowing operations on temporally encoded data accessed from memory.
Temporal computational unit performing tropical algebraic functions
The temporal computational unit includes circuit components arranged to operate on time-delayed wavefronts implementing tropical algebraic functions, specifically vector matrix multiplication (VMM).
The claims collectively cover a temporal state machine architecture combining temporal memory and computation units electrically connected to process time-encoded wavefronts, capable of executing temporal and algebraic operations such as tropical algebra including vector matrix multiplication.
Stated Advantages
Significant efficiency gains due to encoding multi-bit information in single edges rather than multiple digital transitions.
Energy costs of memory read operations become invariant of stored device state, enabling use of high conductance regimes in memristor-based temporal memories with improved linearity and repeatability.
Temporal memories allow lifting causality and invariance constraints of pure race logic, enabling stateful computations and complex architectures beyond one-shot circuits.
Native time-domain encoding and operation reduce overhead from converting between analog, digital, and temporal domains, improving energy efficiency and speed.
Scalable architectures using memristor crossbars and CMOS allow for sub-50 ns timing operations with reasonable bit precision and low energy consumption per line.
Temporal state machines enable systematic composition of temporal computations, supporting higher-order algorithms like graph traversal, DNA alignment, and shortest path computations with favorable edge traversal rates and energy efficiency.
Documented Applications
Implementation of computational circuits and architectures performing operations such as min, max, inhibit, argmin, binarization, and tropical vector matrix multiplication in the temporal domain.
Graph traversal and analytics including execution of shortest path algorithms like Dijkstra's algorithm within temporal state machines using tropical algebraic functions.
Bioinformatics applications such as DNA sequence alignment employing algorithms like Needleman-Wunsch implemented temporally in state machines.
Temporal computing for path planning, navigation, routing, logistics, network and social graph analysis leveraging tropical algebraic time-domain computations.
Use of temporal memory and state machines as foundational components for general-purpose temporal computational frameworks and accelerators.
Interested in licensing this patent?