Good news! Amazing stuff!
"In-memory computing can execute a large vector-matrix multiplication (VMM) within one computing cycle, but the relatively low precision of these analog methods has precluded its use with conventional computing. Song et al. show that for one-memristor-one-transistor arrays, most of the VMM can be performed to arbitrarily high precision before being output as a digital result (see the Perspective by Aimone and Agarwal). After a computation step by an array, subarrays dynamically compensate for residual errors of the previously programmed array. This method was used experimentally to solve partial differential equations with remarkable precision (<10–15 error) and with higher energy efficiency than digital computing.
"
From the perspective abstract:
"In the past 20 years, microelectronics technology has seen a tapering off of Moore's law (the exponential growth of transistor density on computer chips) and an end to Dennard scaling (wherein smaller transistors use correspondingly less power). To create computing systems that are more powerful, faster, and energy efficient for artificial intelligence (AI) and numerical computing applications, new approaches to computing are needed. Attaining these new capabilities not only requires exploration of new materials and devices, but as reported by Song et al. on page 903 of this issue (1), it also necessitates innovative circuits and new mathematical approaches."
From the editor's summary and abstract:
"Editor’s summary
In-memory computing can execute a large vector-matrix multiplication (VMM) within one computing cycle, but the relatively low precision of these analog methods has precluded its use with conventional computing. Song et al. show that for one-memristor-one-transistor arrays, most of the VMM can be performed to arbitrarily high precision before being output as a digital result ... After a computation step by an array, subarrays dynamically compensate for residual errors of the previously programmed array. This method was used experimentally to solve partial differential equations with remarkable precision (<10−15 error) and with higher energy efficiency than digital computing. ...
Abstract
In-memory computing represents an effective method for modeling complex physical systems that are typically challenging for conventional computing architectures but has been hindered by issues such as reading noise and writing variability that restrict scalability, accuracy, and precision in high-performance computations. We propose and demonstrate a circuit architecture and programming protocol that converts the analog computing result to digital at the last step and enables low-precision analog devices to perform high-precision computing. We use a weighted sum of multiple devices to represent one number, in which subsequently programmed devices are used to compensate for preceding programming errors. With a memristor system-on-chip, we experimentally demonstrate high-precision solutions for multiple scientific computing tasks while maintaining a substantial power efficiency advantage over conventional digital approaches."
Overcoming the noise in neural computing (no public access) Circuit strategies can enable noisy analog hardware to achieve high precision
Programming memristor arrays with arbitrarily high precision for analog computing (no public access)
No comments:
Post a Comment