Chinese search giant Baidu's AI chip Kunlun 1 has been mass-produced more than 20,000 pieces and has achieved application deployment, Baidu Chief Technology Officer Wang Haifeng revealed at the 2020 Baidu Cloud Summit held today, adding Kunlun 2 is expected to be mass-produced in the first half of next year.
"Baidu Kunlun" is China's first cloud-based full-featured AI chip developed by Baidu.
Since 2011, Baidu has been developing AI gas pedals based on FPGAs for deep learning computing and started using GPUs at the same time.
In 2018, Baidu announced the launch of its cloud-based full-featured AI chip, "Baidu Kunlun".
Baidu Kunlun 2 uses a 7nm process and offers a 3x performance improvement over Baidu Kunlun 1.
Baidu AI chip boasts up to 3 times the performance of Nvidia
It was previously reported that the Baidu Kunlun chip can achieve actual performance of over 200 TOPS. In comparison, the industry's most advanced Intel T4 GPU can achieve real-world TOPS of around 70.
In the image segmentation YOLOV3 algorithm, Baidu Kunlun's performance is about twice that of T4.
As of today, the Baidu Kunlun chip fully supports speech, image, and NLP models, flexible support for user-defined models, and support for traditional machine learning models.
The chip can achieve complete adaptation with China-made processors, operating systems, and domestic frameworks.
Baidu Kunlun AI processor adapts to China-made Phytium 64-core CPU