Dynamic Differential-Reference TDC Readout for CIM Signal Margin

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The challenge in Computing-In-Memory (CIM) applications is the restricted signal margin due to power supply voltage VDD and high power consumption by readout circuits, leading to errors and inefficiencies in MAC operations.

Innovation Solution

A dynamic differential-reference time-to-digital converter with a configurable main-reference selector and multiple time-to-digital converters, adjusting bias and differential multiplication-and-accumulation values based on a bias reference and setting parameter, to improve signal margin and reduce power consumption.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If multiple word lines are turned on simultaneously to increase bandwidth and reduce power consumption per operation, then computing throughput is improved, but signal margin is restricted causing sensing errors

Engineering Contradiction:
Improvecomputing throughputVSAvoidsensing accuracy
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent changes the parameter of reference voltage from static to dynamic, adjusting it in real-time to match the accumulated MAC values across multiple activated word lines. This dynamic adjustment maintains optimal signal margin even when multiple word lines are simultaneously active, thereby preserving sensing accuracy while enabling high throughput computing operations

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If readout circuits are designed for high accuracy MAC value sensing, then measurement precision is improved, but power consumption increases

Engineering Contradiction:
ImproveMAC value sensing accuracyVSAvoidreadout circuit power consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent implements dynamic reference voltage adjustment in the readout circuit, allowing the reference level to adapt to the actual accumulated MAC values. This dynamic approach enables the circuit to maintain high measurement precision only when necessary, reducing power consumption compared to static high-precision designs that continuously operate at maximum accuracy levels

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

By changing the reference voltage parameter dynamically based on the accumulated signal levels from multiple word line operations, the readout circuit achieves optimal sensing accuracy with reduced power consumption, avoiding the need for continuously high-power operation

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentUS12386592B2Memory array structure with dynamic differential-reference based readout scheme for computing-in-memory applications, dynamic differential-reference time-to-digital converter for computing-in-memory applications and computing method thereof
Publication Date: 2025.08.12 NATIONAL TSING HUA UNIVERSITY
  • US12386592B2 patent drawing
  • US12386592B2 patent drawing
  • US12386592B2 patent drawing

AI summary

A dynamic differential-reference time-to-digital converter for computing-in-memory applications is controlled by a bias reference and a predetermined setting parameter, and includes a configurable main-reference selector and a plurality of time-to-digital converters. The configurable main-reference selector is configured to receive a plurality of edge-output signals, select one of the edge-output signals as a main reference and select others of the edge-output signals as a plurality of edge selected signals according to the predetermined setting parameter. One of the time-to-digital converters is configured to compare the bias reference with the main reference to output a bias multiplication-and-accumulation value, and others of the time-to-digital converters are configured to compare the main reference with the edge selected signals to output a plurality of differential multiplication-and-accumulation values. The bias multiplication-and-accumulation value and the differential multiplication-and-accumulation values are dynamically adjusted according to the bias reference and the predetermined setting parameter.