Memory Centric Artificial Intelligence Processor
|
정홍식 UNIST Nowadays AI takes center stage in science and engineering related to not only IT, but all other manufacturing and health care industries. From hardware point of view, AI-oriented approaches have dragged a lot of attention back to the importance of efficient computing. Until now, Moore’ law has led the IT technologies based on computing performance and cost successfully. However, this trend currently encounters a new challenge both in technical and economical aspects. Therefore, a new type of processors, so-called NPU (Neural Processing Unit), have been developed in various ways to support efficient computing for AI application. There are two main streams in NPU platforms, computer-science-oriented and neuroscience-oriented which have the fundamental differences in their formulations and coding schemes. In any NPU platforms, one of the most important key elements is the synapse which is related to weights in computer-science-oriented and time series of spikes in neuroscience-oriented. The synapse devices have been developed on a basis of memory device technologies which are closely related to SRAM, DRAM, Flash and memristor devices. Currently commercial synapse devices are mainly realized in SRAM and DRAM as weight storage, of which devices have the weakness of volatility. Thus, these devices additionally require external storage devices like Flash memory, which results in frequent data movement and inefficient computing. To overcome this problem, memristor based devices having non-volatility have been developed as 2nd phase synaptic devices for weight storage, of which approach is much more efficient for computing. Also, 3rd phase synaptic devices which achieve weight storage and Matrix-vector multiplication simultaneously have been investigated by using memristor based synaptic devices for the most efficient computing in computer-science-oriented AI hardware platform. Moreover, synapses for neuroscience-oriented platform have been studied by using these type of memory devices. As mentioned, memory devices are the fundamental key factor for AI processors, covering from just weight storage to neuroscience based synapses. Indeed, AI processors heavily rely on synapses which are based on memory device technologies. In this presentation, I will discuss the technology and characteristics of the synaptic devices for various AI computing platforms. Finally, I will suggest the importance of memory centric artificial intelligence processor, especially in Korea. |