Researchers at IBM Research Europe have developed a new mixed-signal in-memory computing chip that could significantly enhance the performance of deep neural networks. The chip, presented in a paper published in Nature Electronics, combines phase-change memory devices with digital computing processors to support the computations of complex Deep Learning Algorithms.

The 64-core chip integrates analog phase-change memory (PCM) cores with digital processing units, connected via an on-chip digital communication network. Each core contains a crossbar array of synaptic unit cells and time-based analog-to-digital converters to transition between the analog and digital domains. The chip also features a global digital processing unit that implements long-short term memory (LSTM) network operations.

One notable advantage of this chip is its ability to perform all computations associated with individual layers of a neural network on-chip, resulting in reduced computation times and power consumption. In testing, the chip achieved remarkable accuracy, with deep neural networks trained on the CIFAR-10 image dataset achieving an accuracy of 92.81%.

This work represents a significant advancement in the development of analog in-memory computing (AIMC) chips that can effectively support the demands of deep learning algorithms. The researchers envision further improvements in performance by combining analog in-memory computing tiles with a mix of specialized digital compute-cores connected through a 2D mesh.

Further research and development in this area could lead to more efficient and powerful AI inference accelerators, enabling advancements in machine learning and artificial intelligence applications.

Source: [Author’s knowledge]

The post Analog In-Memory Computing Chip Shows Promise for Deep Learning Algorithms appeared first on TS2 SPACE.