Adaptive Microbump Data Transmission for ML Memory Bandwidth

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional memory sub-systems face challenges in transmitting data for machine learning operations due to a limited number of I/O pins, which results in insufficient bandwidth and speed, leading to performance delays.

Innovation Solution

The use of microbumps, which are tiny metal bumps that can carry data signals, allows for increased data transmission by providing a larger number of interconnects than traditional I/O pins, and the memory sub-system selectively uses microbumps based on operation conditions to optimize signal integrity and speed.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If traditional I/O pins are used for data transmission, then device complexity is reduced, but data transmission bandwidth and speed are insufficient

Engineering Contradiction:
Improvedata transmission bandwidthVSAvoidinterconnect structure complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent segments the data transmission function across multiple microbumps instead of using a single I/O pin. Each microbump acts as an independent interconnect, allowing parallel data transmission and significantly increasing overall bandwidth while maintaining manageable complexity through modular organization

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transitions from traditional planar I/O pin arrangements to a three-dimensional microbump array configuration. This dimensional change enables denser interconnect packaging and higher data transmission capacity within the same footprint area

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Productivity

If more I/O pins are used to increase data transmission capacity, then bandwidth is improved, but signal integrity deteriorates

Engineering Contradiction:
Improvedata transmission capacityVSAvoidsignal integrity
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent applies different characteristics to different microbumps based on their specific functions. Some microbumps are optimized for high-speed data transmission while others are dedicated to control signals or power, allowing each interconnect to be tuned for its specific role and maintaining signal integrity across the entire array

Inventive Principle:
Principle #3Local quality

3Speed

If microbumps are used to increase data transmission speed, then processing time is reduced, but manufacturing precision requirements increase

Engineering Contradiction:
Improvedata transmission speedVSAvoidmicrobump alignment precision
Core Design Contradiction:
SpeedVSManufacturing precision

Solution Approach 1:

The patent incorporates self-alignment features in the microbump structure design, where the microbumps automatically position themselves relative to corresponding pads during the bonding process. This self-alignment mechanism reduces dependency on high-precision external alignment and simplifies manufacturing while maintaining high data transmission speeds

Inventive Principle:
Principle #25Self-service

Data Source

PatentUS12189505B2Transmission of data for a machine learning operation using different microbumps
Publication Date: 2025.01.07 MICRON TECHNOLOGY INC
  • US12189505B2 patent drawing
  • US12189505B2 patent drawing
  • US12189505B2 patent drawing

AI summary

A system includes a memory device with microbumps and a processing device. The processing device is operatively coupled with the memory device to perform operations. The operations include receiving information indicating a current condition of the machine learning operation while data for the machine learning operation is being transmitted using a first set of microbumps of the plurality of microbumps. Furthermore, the operations include, in response to a change in the condition of the machine learning operation, transmitting subsequent data using a second set of microbumps of the plurality of microbumps, wherein a number of microbumps included in the second set of microbumps is based on the received information.