Graphic processing unit (GPU), data processing unit (DPU), neural network processing unit (NPU)… . It is a system semiconductor that emerged as data to be processed increased exponentially with the development of artificial intelligence (AI) technology.
Samsung Electronics approached it a little differently. It has developed an innovative technology that reduces the role of the central processing unit (CPU) by mounting an AI processor on the memory semiconductor. It is’HBM (high bandwidth memory)-PIM (processor in memory)’ announced by Samsung Electronics on the 17th. Recently, related papers were also published at ISSCC, the most prestigious academic society in the semiconductor field.
The HBM-PIM introduced this time plays the same role as the existing DRAM. In general, a computer operates in a’phone Neumann’ structure in which the CPU retrieves data from the DRAM, operates it, and stores it in a storage device. As data travels between the CPU and DRAM, a bottleneck occurs when the data to be processed increases rapidly. In the field of AI dealing with big data, it has been cited as the reason for the slowing down of computational processing speed.
In order to solve this limitation, Samsung Electronics started to develop a technology that processes data operations in DRAM in advance and transfers them to the CPU. To this end, an AI processor was installed on the 2nd generation HBM’Aquabolt’ developed in 2018. HBM is a memory semiconductor for supercomputers with a higher capacity and much higher speed than DRAM. At 2.4Gb per second (gigabit), 61 full HD movies can be transmitted in 1 second.
The semiconductor that upgraded this is HBM-PIM. Samsung Electronics installed AI engines in each bank (storage logic unit) inside HBM to maximize parallel processing. Compared to the existing 2nd generation HBM, the performance is more than doubled and the system energy is reduced by more than 70%.
As it supports the existing HBM interface, it is also an advantage that performance can be improved without changing hardware or software. In the first half of the year, Samsung Electronics plans to complete the test by mounting HBM-PIM in the AI accelerators of customers.
In the semiconductor industry, it is evaluated that Samsung Electronics has prepared an opportunity to lead the AI semiconductor market. Professor Park Jae-geun of the Department of Convergence Electronics Engineering at Hanyang University said, “The only companies that can design and mass-produce HBM in the world are Samsung Electronics and SK Hynix.”
Samsung Electronics is continuously developing technologies that can perform calculations with AI processors while increasing memory semiconductor capacity and performance.
Reporter Lee Soo-bin [email protected]
Ⓒ Hankyung.com prohibits unauthorized reproduction and redistribution