SK Telecom's AI Solution Accelerates Deep Learning Operations

SK Telecom is turning into an artificial intelligence (AI) company. The company unveiled on June 21 an artificial intelligence (AI) accelerator, which handles AI deep learning more efficiently.

The company's AI acceleration solution, dubbed AI Inference Accelerator (AIX), combines hardware loaded with a neural processing unit (NPU) and a high bandwidth memory (HBM), and software related to an AI algorithm and the voice recognition service Nugu.

SK Telecom expects that the AI accelerator, which enhances the speed and amount of AI processing, will be able to increase the capacity of its AI service Nugu about five times.

SK Telecom's AI acceleration solution is loaded into the accelerator in the form of a small palm-sized card. The accelerator raises the speed of deep running operations up to 20 times when mounted on existing AI servers at data centers.

SK Telecom is expanding the Nugu service to speakers, kids mobile phones and set-top boxes among others.

"We will upgrade our own AI acceleration solution to 2.0 next year and 3.0 in 2020, and finally launch an 'interactive knowledge AI service,'" said Chung Moo-kyong, head of Machine Learning (ML) Infra Lab at SK Telecom’s ICT Technology Center.

“Nugu has to send voice commands to a data center and receive messages from it. When 5G is introduced, the latency problem in this section will be resolved, high-quality services will become possible and furthermore, universal AI services will be provided via various devices."

Meanwhile, artificial intelligence accelerators are a field which global IT companies such as Google, Intel, and Huawei competitively entered. In particular, Nvidia's share price shot up thanks to the parallel processing capability of a graphics processing unit (GPU) which belongs to the company’s main business areas as the capability drew much attention as a key technology in AI and autonomous navigation research. This is because the characteristics of GPUs that operate in parallel are similar to those of AI operations. "An increase in the number and type of AI services is fueling demand for and interest in an NPU, an AI-only processing unit rather than a GPU," an SK Telecom official said. “An NPU-based solution is 16 times more energy efficient than a GPU-based one."