AI Hardware Accelerator System-Level Testing via Segmented Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional test generation techniques are inadequate for system-level testing of AI hardware accelerators due to excessive time and space requirements, making effective test coverage and error analysis impractical.
Innovation Solution
The method involves replicating input data to generate multiple identical units for hardware execution, simulating input data for software prediction, and configuring a daisy chain of AI primitives to reduce test generation and execution times, enabling efficient validation and error analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional test generation techniques are used for system-level testing of AI hardware accelerators, then test coverage can be achieved, but test generation time and space requirements become prohibitively large
Solution Approach 1:
The patent segments the testing process into two distinct phases: (1) a fast first pass that executes test cases and generates pass/fail results without detailed analysis, and (2) a second pass that performs comprehensive error analysis only on failed test cases. This segmentation allows the system to achieve complete test coverage while minimizing the time spent on analysis, as the expensive error analysis is performed selectively only when needed.
Solution Approach 2:
The patent applies partial action by performing full error analysis only on a subset of test cases (those that fail) rather than on all test cases. The first pass performs minimal action (just execution and pass/fail determination), while the second pass performs excessive action (detailed error analysis) only where necessary. This selective application of analysis resources resolves the contradiction between achieving complete coverage and minimizing total analysis time.
2Reliability
If conventional test generation techniques are used for system-level testing of AI hardware accelerators, then test coverage can be achieved, but test space requirements become prohibitively large
Solution Approach 1:
The patent extracts and separates the error analysis function from the main test execution flow. By taking out the detailed error analysis step and performing it only on failed test cases in a second pass, the system avoids generating and storing large amounts of test data for all test cases. Only the minimal pass/fail results from the first pass need to be stored, significantly reducing the test space requirements while maintaining complete coverage.
Solution Approach 2:
The testing process is segmented into two phases with different data storage requirements. The first pass generates minimal result data (pass/fail status), while the second pass generates detailed error analysis data only for failed cases. This segmentation reduces the overall test data volume that must be stored and processed, resolving the contradiction between achieving complete coverage and minimizing test space requirements.
3Productivity
If conventional test generation techniques are used, then testing can be performed, but result prediction time becomes prohibitive
Solution Approach 1:
The patent performs preliminary action by executing all test cases in a first pass to establish baseline pass/fail results before performing any detailed error analysis. This preliminary execution phase allows the system to predict which test cases will fail and schedule the expensive error analysis only for those cases, significantly reducing the total result prediction time compared to analyzing all test cases in detail regardless of their expected outcome.
Solution Approach 2:
The patent applies partial action by performing detailed error analysis only on a subset of test cases (those that fail in the first pass) rather than on all test cases. This selective analysis approach maintains high testing efficiency by avoiding redundant analysis of passing test cases while still providing comprehensive error analysis where needed, thus resolving the contradiction between productivity and prediction time.
4Difficulty of detecting and measuring
If conventional test generation techniques are used, then testing can be performed, but error analysis and printing become impossible due to test data size
Solution Approach 1:
The patent extracts the error analysis function from the main test execution flow and performs it separately in a second pass only on failed test cases. This extraction allows comprehensive error analysis and printing to be performed on a manageable subset of data (only failed cases) rather than attempting to analyze all test data at once, making error analysis and printing feasible despite the overall test data volume.
Solution Approach 2:
The patent applies partial action by performing comprehensive error analysis only on the subset of test cases that fail, rather than attempting to analyze all test cases. This selective approach makes error analysis and printing possible by focusing computational resources on the relevant subset of data, resolving the contradiction between error analysis capability and test data size constraints.
Data Source
AI summary
Methods, system, and computer program product are disclosed for enhanced system level testing of an AI hardware accelerator comprising Artificial Intelligence (AI) primitives. The system level testing provides uncompromised test coverage of the AI hardware accelerator while enabling efficient and effective testing performance. Enhanced test data generation, result prediction, and test execution are provided for system level testing of AI hardware accelerator including the AI primitives.