In desktops, laptops and smartphones, computing modules and memory can be allocated. This approach is known as "von Neumann architecture" - after the scientist John von Neumann, one of the pioneers in the field of digital computing. In such an architecture, data is constantly moving between memory and the computing device, which is slow and not very efficient.
All-in-one computing
The solution to this problem can be "computational memory" - a technology also known as "in-memory computing." In this case, only the physical properties of computer memory are used for storing and processing information.
A research team at IBM announced a breakthrough achievement in computational memory by successfully executing a machine learning algorithm on a group of 1 million phase change memory devices (EPF).
The ESF device was created from a germanium-antimony-telluride alloy sandwiched between a pair of electrodes. This prototype technology is expected to deliver 200x improvements in speed and energy efficiency.
Suitable for artificial intelligence
Promotional video:
The ESP device performs calculations using the crystallization mechanism - when an electric current passes, the disordered arrangement of atoms changes to an ordered, that is, crystalline. Scientists demonstrated ESF technology with two time-controlled examples and compared it with traditional machine learning methods.
The ability to perform calculations faster will affect the overall performance of computers. For IBM, this means more power in artificial intelligence applications.
CMOS technology has reached its limit, and in order to overcome its limitations, a radical change in the processor-memory paradigm is inevitable.
Computational memory expands the ability to process data in real time, which is very important now that companies are focusing on analytical information processing. Industry giants such as Amazon and Google are making AI the centerpiece of their business, so the speed of AI computers is in high demand.
Vadim Tarabarko