Riassunto analitico
Artificial Intelligence (AI) is driving innovation across various fields and becoming essential in everyday life. However, the rapid growth of this technology in recent years has placed significant strain on both the energy and semiconductor industries. To meet the demand for efficient computation, analog electronics, particularly crossbar architectures with processing-in-memory capabilities, present a promising solution. These architectures exploit emerging non-volatile memories to enable efficient execution of matrix-vector multiplications, a key operation in modern AI algorithms.
However, the effect of line parasitic resistance of crossbar structures deteriorates analog output signals, leading to a loss of computational accuracy. This problem creates a critical need for models that can provide fast and accurate estimations of these distortions to enable effective compensation strategies, such as parasitic-aware training of artificial neural networks (ANN). While numerous models have been proposed in the literature, a comprehensive performance benchmarking is still missing.
In this thesis, an open-source and flexible simulation framework for benchmarking crossbar line parasitic resistance models was developed and used to benchmark and compare the performance of several models from the literature. Initial simulations on binarized crossbar arrays under various operating conditions reveal a trade-off between analytical and iterative models. Analytical models offer computational efficiency but struggle under challenging conditions, such as high (i.e., > 4Ω) parasitic resistance and large array sizes (i.e., > 64x64). In contrast, iterative models provide greater accuracy but require significantly longer execution time. In fact, simulations showed that the error of analytical models remains below a reasonable level (i.e., 5%) only for array sizes smaller than 64x64 and with interconnect resistances below 1Ω. These findings are further confirmed by the results of a parasitic-aware training task performed for a simple classifier for the MNIST dataset. All the crossbar parasitic resistance models were able to achieve trained networks with over 97% accuracy on the test set, but only when fully connected layers were mapped onto crossbar tiles smaller than 64×64. As tile size increases, accuracy deteriorates rapidly, making iterative models the only practical option despite their significant overhead in training time and the high number of epochs required.
These insights provide a valuable guide for researchers and engineers in selecting optimal modeling techniques for crossbar-based accelerator design. From the results presented in this thesis, future directions can be outlined to enhance performance, including the formulation of better models and the development of more effective weight mapping strategies.
|
