Samsung working on in-memory computing MRAM

Samsung has managed to equip MRAM technology with the ability to also perform arithmetic operations within the memory. This means that three different chips are now combined in one architecture. Normally, computers consist of the processor, the main memory and a permanent memory in their basic structure.

Combinations of two variants have already been implemented in various ways: Computing processes in main memory are no longer unknown and the combination of main memory and fixed memory in technologies such as magnetoresistive random access memory (MRAM) has also been promoted for some time. Now the combination of all three models in one chip has also succeeded, as Samsung announced. According to a report by the industry journal, the engineers at the South Korean group have EETimes MRAM components equipped with the ability to perform calculations within the memory architecture. This in-memory computing can bring high gains in efficiency in some application areas.

Best AI training

Of course, the focus here is on tasks in which relatively simple arithmetic operations on large amounts of data have to be mastered. This avoids moving large amounts of information from the hard drive to the main memory and from there to the CPU and back again, which ultimately saves a lot of time and energy. Of course, this primarily affects the training of AI systems, in which huge amounts of data have to be evaluated using statistical methods in order to teach the algorithm the required patterns.

Tests with smaller MRAM units of this type, for example, achieved real-time accuracies of 98 and 93 percent for handwritten number recognition and face recognition, respectively. The Samsung engineers emphasize that this only affects the first generation of the new architecture and they are optimistic that the chip design can still be significantly improved.