Samsung Electronics announced on the 17th that it has succeeded in developing an intelligent memory semiconductor that can simultaneously process data storage and operations for the first time in the world. Memory semiconductors that could only store data are equipped with an artificial intelligence (AI) engine to self-process some computational functions, and the industry expects to change the paradigm of memory semiconductors.
In general, semiconductors are divided into memory semiconductors in charge of storage functions and system semiconductors in operation functions.
In the existing computer structure, called the von Neumann structure, when a user inputs a task necessary for an electronic device, system semiconductors such as CPU (central processing unit) fetch and execute instructions from the memory semiconductor, and store the result back in the memory device. Work sequentially. In this process, when the amount of data exchanged between the system semiconductor and the memory semiconductor increases, the processing of the work is delayed (the von Neumann bottleneck).
In particular, as the use of artificial intelligence technology and large-capacity data centers has recently increased, the amount of data to be processed has increased exponentially, and this problem is expected to become more and more prominent.
Accordingly, Samsung Electronics installed an artificial intelligence processor on the memory semiconductor based on the process know-how accumulated in the DRAM field. As memory semiconductors can process some operations, the amount of data that needs to be exchanged with system semiconductors can be reduced. If the amount of data movement between memory and system semiconductors decreases, the data bottleneck is expected to be resolved.
Samsung Electronics recently unveiled HBM-PIM technology through a paper at ISSCC, one of the world’s most prestigious academic conferences in the semiconductor field. HBM stands for High Bandwidth Memory, and PIM (Processing-in-Memory) is a next-generation new concept convergence technology that adds processor functions necessary for computational tasks inside the memory.
According to Samsung Electronics, when HBM-PIM is applied to the existing artificial intelligence system, the performance is more than doubled compared to the previous HBM2. In addition, the energy consumed by the system is reduced by more than 70%.
In the industry, PIM technology is expected to be widely used in AI semiconductors and data centers, which are the next-generation semiconductor markets. This is because processing and analysis of enormous amounts of data is required to increase the utilization of AI technology.
In addition, in the server semiconductor market where the demand is rapidly increasing recently, the fact that power consumption is emerging as a major factor of consideration also brightens the outlook. This is because the power consumption of a data center that manages 400,000 servers is equivalent to the total power consumption of a city with a population of 200,000. In particular, it is known that about 40% of the power consumption of this data center is used in main memory.
[박재영 기자]
[ⓒ 매일경제 & mk.co.kr, 무단전재 및 재배포 금지]