Functional Test Generation for AI Accelerators using Bayesian Optimization
Unmet Need
Commercial use of artificial intelligence (AI) accelerator technology has grown rapidly as major advances are made in the field. However, when AI accelerators run into process-halting faults, substantial computational effort and time is required to test for and uncover fault inducing sources, such as faulty neural nodes that propagate errors into subsequent calculations. There is a need for enhancing AI inferencing accelerator systems with new machine learning approaches that are capable of efficiently and effectively seeking out faults through sophisticated optimization and test generation protocols with a reduction in computational expense.
Technology
Duke inventors have developed a black-box optimization method to generate functional test patterns for AI inferencing accelerators used in various commercial applications. This is intended to be an optimization method that is deployed within AI accelerator systems to provide a faster and less computationally intense approach for fault detection during inference processes. Specifically, this approach utilizes Bayesian optimization for targeted test-image generation to overcome stuck-at faults in systolic array-based AI accelerators with features such as test-pattern compaction and various error regularization methods. Currently, this technology has been demonstrated as a fully developed optimization method for functional test generation to achieve high fault coverage. This is accomplished using only a small set of test images for faults in 16-bit and 32-bit floating point processing elements of the systolic array in AI accelerators.
Other Applications
This technology can be deployed in essentially any AI acceleration scheme as an effective optimization and test generation tool.
Advantages
- Faster fault-searching
- Reduces computational expense
- More effective test pattern generation
- Reduces “over-testing”