AI Chip Based on Neural Network Processing Units
The Korea Electronics and Telecommunications Research Institute (ETRI) and SK Telecom have jointly developed an artificial intelligence (AI) semiconductor based on neural network processing units (NPUs), which can be applied to data centers and Internet of Things devices, the Ministry of Science and ICT announced on April 7.
ETRI and SK Telecom developed Korea’s first ultra-low-power AI semiconductor (AB9) for servers that can be applied to high-performance servers at cloud data centers. Currently, the CPUs and GPUs used for AI computation consume a large amount of power and have a large chip size, which limits their efficiency.
The research team realized high computational performance by integrating 16,384 computing devices in an area of the size of a coin (17 mm wide and 23 mm long). In particular, it is possible to minimize power consumption by applying software technology that can operate and cut off power to each computing device.
The new AI semiconductor can process data 40 trillion times per second and consumes power as low as 15W to 40 W. The research team will conduct empirical tests at SK Telecom’s data centers that provide intelligent CCTV and voice recognition services in the second half of the year.
ETRI has also developed a visual intelligence chip (VIC) that recognizes vision through AI, in cooperation with the Korea Electronics Technology Institute and fabless companies.
An AI semiconductor can recognize objects 30 times per second on a circuit area equivalent to half the size of an adult's nail (5 mm wide and 5 mm long) so it can be used in mobile and IoT devices. Its power consumption is only 0.5 W, which is one-tenth of conventional semiconductors’ power consumption.
The research team plans to promote demonstration and commercialization by applying it to intelligent CCTVs and drones in the second half.