High-Bandwidth Memory Thermal Control Using Localized DDVFS
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing monolithic thermal throttling techniques in high-bandwidth memory (HBM) systems inefficiently cool majority portions of core dies due to treating the HBM as a monolith, despite minority portions exceeding permissible temperatures, leading to wasteful performance diminishment.
Innovation Solution
Implementing differentiated dynamic voltage and frequency scaling (DDVFS) that adjusts clock signals and power supply voltages at a granular level, such as bank-wide, group-wide, channel-wide, or core-wide, based on localized temperature and threshold voltage sensing, to efficiently control temperature variations within the HBM.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Temperature
If monolithic thermal throttling is used to cool the HBM, then the temperature of minority portions exceeding permissible temperatures is reduced, but the performance of majority portions is unnecessarily diminished
Solution Approach 1:
The HBM is divided into multiple segments (banks, groups, channels, or core-dies) with independent temperature monitoring and control. Each segment has its own sensing unit that monitors local temperature and triggers throttling only in that specific segment when temperature exceeds the threshold, rather than treating the entire HBM as a single monolithic unit.
Solution Approach 2:
Temperature control is applied locally to specific segments rather than uniformly across the entire HBM. The sensing units are distributed throughout the HBM structure, enabling localized detection and response to temperature conditions in each bank, group, channel, or core-die independently.
2Device complexity
If monolithic thermal throttling treats the HBM as a single unit, then temperature control is simplified, but cooling efficiency is reduced due to unnecessary throttling of non-hot regions
Solution Approach 1:
The HBM is segmented into multiple independently controllable units (banks, groups, channels, or core-dies), each with its own temperature sensing unit. This segmentation enables precise localization of hot spots and application of cooling measures only where needed, improving cooling efficiency without significantly increasing overall system complexity.
Solution Approach 2:
Each segment of the HBM is equipped with its own temperature sensing unit that autonomously monitors its local temperature and triggers throttling control when necessary. This distributed self-monitoring approach eliminates the need for complex centralized temperature management while improving responsiveness and efficiency.
Data Source
AI summary
A system (for controlling temperatures in a memory) includes: a high bandwidth memory (HBM) including core dies, the HBM being arranged into portions, each of the portions including memory cells, the HBM further including: a first sensing unit configured to generate a first environmental signal corresponding to a first transistor in a corresponding first one of the memory cells in a first one of the portions; and a second sensing unit configured to generate a second environmental signal corresponding to a second transistor in a corresponding second one of the memory cells in a second one of the portions; and a controller configured to perform non-monolithic temperature control of the HBM based on the one or more first environmental signals and the one or more second environmental signals such that temperature of one or more of the portions is differentially controlled.