K3 Test system for neuromorphic memory
In project K3, the companies aixACCT systems and AMOtronics are working on new drive concepts and test technologies, both for memristive single cells and for matrix structures. For this purpose, the performance of the aixACCT KI tester is evaluated and further developed to achieve an improvement in the speed and functionality of the hardware. The aixACCT KI tester has an ultra-fast current limiting circuit whose maximum and minimum current can be set via a communication interface. The limiting of the current takes place in the range of a few nanoseconds and effectively protects the tested cells from thermal overload at the switching moment due to low overshoot.
In single-cell studies, switching kinetics is the main component of the research within the overall project. In the K3 project, this involves recording the sample current via the current/voltage converters used with a new 10 gigasample/s analog-to-digital converter to be developed, enabling temporal resolution in the sub-nanosecond range with a bit depth of 16 bits. This includes a comprehensive functional upgrade of the FPGAs at the A/D and D/A converters.
By using the aixACCT AI tester, the analysis of memristive arrays up to a size of 32x32 cells can be performed. The large amounts of obtained measurement data in statistically significant numbers are analyzed via AI-based algorithms developed within the project. Machine-learning methods for reliability analysis, lifetime estimation as well as optimization of production-related parameters will be developed and transferred to industry-relevant scales in order to apply them to future integrated circuits. A crucial step of data analysis here is the early detection of manufacturing-related artifacts that may indicate subsequently occurring defects and accelerated aging of memristive hardware. Pattern recognition is performed in the recorded current-voltage characteristics and is intended to demonstrate the performance of the AI-based algorithms used. The AI methods developed by aixACCT systems can thus be used in the future to optimize new generations of AI chips and enable predictions to be made about the lifetime of selected structures, as well as conclusions to be drawn about relevant process parameters in particular.
