SKT is using Xilinx's Kintex UltraScale FPGAs to run its automatic speech-recognition (ASR) application to accelerate NUGU, its voice-activated assistant. According to SKT, it has been able to achieve up to five times higher performance in ASR applications when compared to GPUs, and more importantly, 16 times better performance-per-watt.
This is the first commercial adoption of FPGA accelerators in the AI domain for large-scale data centres in South Korea.
The FPGA-based accelerator lowers the total cost of ownership by populating existing CPU-only servers with efficient Xilinx FPGA add-in cards. The ASR servers simply accelerate multiple voice service channels with Xilinx FPGA cards in their empty slots. One FPGA card provides more than five times the performance of a single server, resulting in substantial cost savings.
“Over many years we have seen the shape of the industry evolve, and are proud to be at the forefront of developing AI accelerators. By designing our solution based on the Xilinx KCU1500 board and our own bitstream image, we have developed a cost-effective, high-performance application,” said Kang-Won Lee, senior vice president, software research and development centre at SKT.
The adaptive nature of Xilinx FPGAs enables fast deployment of custom hardware accelerators for the rapidly evolving field of AI and deep learning. SK Telecom has joined a fast-growing list of high-profile commercial data centres that have deployed FPGAs for compute acceleration.
