On the morning of July 4th, the 2018 Baidu AI Developers Conference was held in Beijing.
At the meeting, Li Yanhong announced China's first cloud AI full-featured AI chip 'Kunlun', which was independently developed by Baidu, including training chip Kunlun 818-300, reasoning chip Kunlun 818-100.
It is reported that, Kunlun is by far the most powerful AI chip in the industry, providing 260TOPS (260 trillion calculations per second) performance at 100 watts + power consumption. Previously, NVIDIA's Xavier chip for auto L4.L5 autopilot (8-core ARM CPU + 512 CUDA) performance was 30TOPS.
In terms of specifications, The 'Kunlun' chip uses Samsung's 14nm process, 512GB/s memory bandwidth, and consists of tens of thousands of small cores.
Baidu introduced, 'Kunlun' has Efficient (Specially optimized for voice, NLP, images, etc.), low cost (10 times lower cost under equivalent performance) and Easy to use (Supports multiple deep learning frameworks such as paddle; high flexibility of programming; flexible support for training and prediction).
'Kunlun' is a chip that China has spawned in the practice of large-scale AI computing. Based on Baidu's 8-year CPU, GPU and FPGA AI accelerator development, more than 20 iterations will be made, and the future will be for smart cars, smart devices, voice. More scenes such as images are another milestone for Chinese AI chips.
