Samsung presented a report on new developments in the field of MRAM Magnetoresistive Random-Access Memory technology). The company said it has found a way to make in-memory computing practical using this technology.
As a non-volatile memory, MRAM arrays provide data retention up to 10 years at 200°C, withstand up to 106 write cycles and 1012 read cycles. In addition to high reliability, MRAM demonstrates a good level of product recovery (only 0.1% rejection). MRAM technology will overcome the reached limit of miniaturization of memory elements. In addition, it can be operated in a wide range of environments, making it optimal for various IoT devices.
Magneto-resistive random access memory technology has been around for over a decade, and Samsung announced major developments for the first time in almost three years. In a standard computer architecture, data is stored in memory chips, and data calculations are performed in separate processor chips. In contrast, in-memory computing allows both data storage and computation in a memory network. Since this allows the large amount of data stored on the memory network itself to be handled without having to be moved, power consumption is greatly reduced. Thus, memory computing has become one of the promising technologies for the production of next-generation low-power semiconductor chips with artificial intelligence.
The biggest disadvantage of MRAM is that it takes a lot of space for a large memory capacity. This means that the technology is not suitable for most consumer devices, which require several gigabytes of memory to fully function.
Samsung is exploring the use of MRAM in the IoT and AI markets. It is developments in the field of AI that have become the focus of the current announcement of the company. At its core, in-memory computing is how the human brain works, and the emulation of this function – with so-called “neuromorphic” computers like the Intel Loihi – has led to some breakthroughs in AI research.
Neuromorphic computers do not require a large amount of memory due to the way they work. Instead, each synapse in the network has a maximum of a couple of hundred kilobytes of RAM.
Another challenge is the low resistance of MRAM, which prevents it from taking advantage of the reduced power consumption of a standard in-memory computing architecture. Samsung researchers proposed to solve this problem with an architectural innovation. They succeeded in developing an MRAM array chip that demonstrates in-memory computing by replacing the standard "current sum" memory computing architecture with a new "resistance sum" in-memory computing architecture. The company claims to have achieved in-memory computing efficiency by running AI to classify images on MRAM circuits and achieving 98% accuracy for handwritten digit recognition and 93% accuracy for face recognition.
Samsung expects its MRAM to find applications in "next-generation low-power AI chip technologies" and specifically mentioned neuromorphic computing as a possible use case.
In 2020, MIT engineers showed a "brain on a chip". Smaller than a piece of confetti, it is made up of tens of thousands of artificial brain synapses, or memristors, silicon components that mimic how synapses work to transmit information to the human brain. The device was able to process images.
bbabo.Net