NPU Built-In Self-Test for Runtime Defect Isolation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural processing units (NPUs) in critical systems like autonomous vehicles and drones may operate erratically due to undetected minor defects during manufacturing, which are amplified over time, leading to unpredictable AI behavior and potential safety hazards.
Innovation Solution
Implementing a runtime testing system that includes a component tester to perform scan tests on NPUs, connecting flip-flops to form scan chains, applying test inputs, and analyzing operational logic to detect defects, allowing for the isolation or deactivation of faulty components.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If pre-shipment testing is performed on NPU components, then manufacturing defects can be detected, but minor or minute defects may remain undetected and be handed over to users
Solution Approach 1:
The patent implements built-in self-test (BIST) functionality that performs testing before the NPU is shipped from the factory. Test circuits are integrated within the NPU architecture to conduct preliminary defect detection on functional components such as processing elements, memory units, and interconnect structures before deployment, ensuring that only properly functioning components are installed in systems.
Solution Approach 2:
The patent introduces test circuits as intermediary components that facilitate detailed defect detection. These test circuits include test data generators, test pattern generators, and response analyzers that act as mediators between the functional components and the testing system, enabling precise detection of minor defects through specialized test patterns and analysis mechanisms.
2Productivity
If NPU components operate continuously in mission-critical systems, then system functionality is maintained, but undetected defects are amplified over time leading to erroneous operation
Solution Approach 1:
The patent implements periodic health monitoring and self-diagnosis mechanisms that continuously or periodically assess the operational status of NPU components during runtime. Test circuits periodically execute diagnostic routines to detect emerging defects before they amplify into critical failures, maintaining system reliability through continuous monitoring and early defect detection.
Solution Approach 2:
The patent incorporates feedback mechanisms where test results from operational monitoring are fed back to system controllers. When defects are detected during operation, the feedback system triggers alerts, logs errors, or activates redundancy mechanisms, enabling real-time response to maintain operational reliability and prevent erroneous AI operation results.
3Adaptability or versatility
If NPU is used in autonomous driving vehicles, drones, or AI robots, then advanced AI functionality is achieved, but unpredictable AI operation results may occur due to NPU failures
Solution Approach 1:
The patent implements prior cushioning by integrating redundant computational resources and error correction mechanisms within the NPU architecture. Test circuits detect potential failures before they affect AI operations, and redundant components provide backup functionality, cushioning against the harmful effects of defects in safety-critical applications like autonomous vehicles and AI robots.
Solution Approach 2:
The patent enables self-service through built-in self-test and self-diagnosis capabilities that allow the NPU to autonomously detect and report its own defects. The test circuits continuously monitor functional components and automatically identify failures, enabling the system to self-diagnose issues without external intervention and maintain reliable operation in autonomous systems.
Data Source
AI summary
A neural processing unit (NPU) for testing a component during runtime is provided. The NPU may include a plurality of functional components including a first functional component and a second functional component. At least one of the plurality of functional components may be driven for calculation of an artificial neural network. Another one of the plurality of functional components may be selected as a component under test (CUT). A scan test may be performed on the at least one functional component selected as the CUT. A tester for detecting a defect of an NPU is also provided. The tester may include a component tester configured to communicate with at least one functional component of the NPU, select the at least one functional component as a CUT, and perform a scan test for the selected CUT.