July 4, 2018 // 11:23 a.m.
Chinese search and advertising giant Baidu has announced what it claims is the nation's first 'cloud-to-edge AI [artificial intelligence] chip': The Kunlun family.
Joining its western rival Google in the race to develop high-performance yet energy-efficient platforms for artificial intelligence and deep learning projects, Baidu has been working with field-programmable gate array (FPGA) and graphics processing unit (GPU) accelerators since 2011. Its latest creation, though, is a family of dedicated application-specific integrated circuits (ASICs) created specifically for artificial intelligence: Kunlun, which it is to be assumed is not named for the mythical realm under the scatterbrained protection of Marvel's The Immortal Iron Fist.
Built on Samsung's 14nm process, the many-core primary Kunlun chip design is claimed to be 30 times faster than the company's original FPGA-based design at 260 trillion operations per second (TOPS) with a 100W thermal design profile (TDP). The chip is also claimed to have 512GB/s memory throughput performance, though the company has not indicated how much memory it can access.
Baidu has confirmed that while the Kunlun family - which includes dedicated chips for training and inference workloads - are functional now, the company is going to continue to iterate the design with a view to making a splash in intelligent vehicle systems, voice recognition, and image recognition.