The world's first commercial in-memory computing chip has been released, and AI processing has improved energy efficiency by more than 10 times.
CTOnews.com news on December 28, it is well known that the common von Neumann architecture of computers at present consists of two parts: memory units and computing units. Computers need to transfer data between memory units and computing units. If the speed of memory cannot keep up with the speed of CPU computing, there may be a "memory wall".
In-memory computing, which has sprung up in the past two years, is to embed the computing unit into memory, which can effectively overcome the bottleneck of von Neumann architecture. CTOnews.com learned that the technology was first proposed in the 1960s, but it has not been commercialized, and big companies such as Samsung are currently working on it.
Samsung memory computing technology announced: the world's first MRAM, paving the way for the next generation of AI chips.
In March 2022, WTM2101, the first in-memory computing SoC chip in the world, was officially put into the market. Less than a year ago, WTM2101 has successfully achieved commercial end-to-side, providing voice, video and other AI processing solutions and helping products to achieve more than 10 times energy efficiency.
In addition, ▲ Technology will launch WTM-8 series products in 2023 and 2024, which will increase its computing power by 500 to 600 times. Around 2025, it will launch WTM-C series products that can be used in edge servers and so on. With the advancement of integration scale, technology and other aspects, in-memory computing products are expected to increase by 5 to 10 times per year in the next few years.
According to the official website of ▲ Technology Enterprise, Zhixun Technology was founded in 2017, focusing on the field of in-memory computing chips, innovating the use of Flash memory to complete neural network storage and computing, solving the memory wall problem of AI, improving computing efficiency and reducing costs.
The R & D team is set up by Dr. Wang Shaodi and Dr. Guo Xinjie together with a number of scholars and industry practitioners, with an average of more than 10 years of industry working experience, which is comparable to the international level of deposit and accounting and chip technology team.
▲ WTM2101 Chip's WTM2101 chips are adapted to low-power AIoT applications and can perform large-scale deep learning operations using microwatt to milliwatt power consumption, especially suitable for intelligent voice and intelligent health services in wearable devices. Mass production and market applications have been completed. WTM8 series chips are designed for 4-32Tops computing products and are used for real-time processing of 4K-8K video.
