Baidu, a leading Chinese company is now presenting a specially designed AI chip called KunLun.
Baidu ‘Kunlun’ is the first IA chip made in China
Baidu today announced KunLun, the first cloud-based IA chip created in China, which is built to meet the high-performance requirements of a wide variety of scenarios implementing Artificial Intelligence.
KunLun is a high-performance and cost-effective solution for the high processing demands of AI. Take advantage of Baidu’s AI ecosystem, which includes search ranking scenarios and deep learning frameworks such as PaddlePaddle. Baidu’s years of experience in optimizing the performance of these services and frameworks involving IA provided the company with the expertise needed to build a world-class IA chip.
Baidu shows AI chips for training and inferencing
Baidu’s development of an IA processor began in 2011 based on FPGA for deep learning and GPUs have been used in data centers ever since. KunLun, which consists of thousands of small cores, has a computing capacity that is nearly 30 times faster than the original FPGA-based accelerator. Other key specifications include 14nm Samsung engineering, 512GB/second memory bandwidth, as well as 260TOPS while consuming 100 watts of power.
KunLun is a high-performance and cost-effective solution for the high processing demands of AI. Take advantage of Baidu’s AI ecosystem, which includes search ranking scenarios and deep learning frameworks such as PaddlePaddle. Baidu’s years of experience in optimizing the performance of these services and frameworks involving IA provided the company with the expertise needed to build a world-class IA chip.
Baidu’s development of an IA processor began in 2011 based on FPGA for deep learning and GPUs have been used in data centers ever since. KunLun, which consists of thousands of small cores, has a computing capacity that is nearly 30 times faster than the original FPGA-based accelerator. Other key specifications include 14nm Samsung engineering, 512GB/second memory bandwidth, as well as 260TOPS while consuming 100 watts of power.
According to Li, the Kunlun can be used for training, this variant is called 818-300, in addition, the chip is suitable for inferencing, where it runs under the label 818-100. 260 teraops with 512 GByte/s bandwidth at over 100 watts power consumption, without specifying the calculation accuracy, the CEO put the speed at least. For comparison: Google’s TPU v2 creates 45 teraflops with FP32 and probably over 200 watts, 16 bits can be used for multiplication. The water-cooled TPU v3 will deliver 90 teraflops per chip.
Baidu Brain 3.0 as new AI service
Baidu describes the Kunlun in its two variants as suitable for deep learning in general, for speech and word processing or for autonomous vehicles. Baidu also presented the AI bus called Apolong, which will be available from the beginning of 2019 in Chinese cities such as Beijing or in Japanese cities such as Tokyo. The Kunlun chips, which have been under development since 2011, will in future be used for Baidu Brain 3.0, Baidu’s AI service, among others.
The Chinese company did not specify exactly when Baidu will use the Kunlun in its systems. However, the chip will only be the first of a series of ASICs for artificial intelligence, so further implementations are expected.