Self-Repairable AI Chip Tiles for Silent Data Corruption
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Silent data corruption (SDC) in AI/ML accelerator chips, such as TPUs, leads to costly discarding of entire machines due to undetectable errors, impacting data center productivity and customer confidence, with current industry approaches limited to detecting and scrapping defective chips.
Innovation Solution
A self-repairable chip design that detects data corruption, identifies defective tiles, and replaces them with redundant rows or columns, determining critical vs. auxiliary logic units, allowing for on-site repair and maintaining system functionality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If design-for-test checkpoints and SDC screening are performed during manufacturing, then defective chips can be detected and screened out, but the chip complexity and manufacturing cost increase
Solution Approach 1:
The chip is divided into multiple independent processing tiles, each capable of being individually identified and replaced. This segmentation allows localized repair rather than whole-chip replacement, reducing the impact of defects while maintaining overall system functionality.
Solution Approach 2:
Redundant processing tiles are pre-configured in the chip architecture before deployment. When a defect is detected, the system can immediately switch to using the redundant tile without requiring complex real-time reconfiguration or external intervention, thus maintaining reliability while managing complexity.
2Reliability
If entire machines with 4 to 8 AI/ML chips are discarded when one chip is defective, then SDC errors are eliminated, but material and logistics costs increase significantly
Solution Approach 1:
The system treats each processing tile as an independent replaceable unit rather than requiring replacement of entire chips or machines. This granular approach allows repair of only the specific defective component, dramatically reducing material loss and replacement costs.
Solution Approach 2:
The chip includes self-diagnostic capabilities that automatically identify defective tiles and trigger replacement procedures without external intervention. This self-service approach enables rapid, targeted repairs that eliminate SDC errors while minimizing resource waste.
3Ease of repair
If redundant rows or columns are added to processing units, then defective tiles can be replaced, but the chip area and manufacturing complexity increase
Solution Approach 1:
Redundant processing tiles are strategically placed only where needed to cover potential defect locations, rather than duplicating the entire chip structure. This localized redundancy provides repair capability while minimizing the additional area required.
Solution Approach 2:
The chip includes a limited number of redundant tiles that provide sufficient repair capability for the most common defect patterns, without over-provisioning redundancy. This partial redundancy approach balances repair ease with area constraints.
Data Source
AI summary
The disclosure provides a self-repairable chip for silent data corruption (SDC) issues. The self-repairable chip is configured to detect data corruption in a plurality of processing units, identifying the defective tiles of the processing units, and repairing the defective tiles by replacing the defective tiles with redundant rows or columns. The self-repairable chip is further configured to determine whether one or more logics where the data corruption occurred must work or may work and replace the defective tiles with redundant rows or columns when the logic where the data corruption occurred may work.