Baidu Kunlun AI chip is adapting to the China-made Feiteng server for performance tuning, Ouyang Jian, general manager of Baidu Kunlun Chip, said at the "2019 Feiteng Ecological Partner Conference" held on Thursday.
Kunlun AI chip is based on Baidu's self-developed XPU neural processor architecture. It uses Samsung's 2.5D I-Cube packaging process, connects the SoC main chip and two HBM2 high-bandwidth memories through an interposer, and is packaged on a substrate.
It provides up to 512GB / s memory bandwidth and supports PCIe 4.0 x8, which can provide 260TOPS (260 trillion operations per second) computing power at 150W power consumption. The chip is expected to be in mass production early next year.
Ouyang Jian introduced that the adaptation to the Feiteng server includes three aspects: computing power, architecture, and technology.
In terms of new computing power, AI's demand for computing power is endless. Feiteng itself is a high-performance 64-bit processor. Kunlun chips are based on Feiteng, which can bring new computing power.
The China-made Feiteng CPU general server was officially introduced by Inspur Group in 2016. At the new architecture level, Feiteng is a universal computing architecture from the cloud to the end.
The same architecture supports different scenarios. Baidu Kunlun uses the XPU architecture. The combination of the two has the opportunity to create a new architecture. Baidu Kunlun calls it a "universal architecture."
At the new technology level, the iteration of the information industry, and underlying technologies such as chips and operating systems, have always been dominated by foreign giants. But each iteration is an opportunity to shuffle. Currently in the era of AI + intelligent cloud, the chip architecture has changed a lot, and many new operating systems have emerged.
Baidu Kunlun + OS made in China, plus the framework of deep learning made in China, coupled with the application of the new era, new technologies will emerge, and this technology will be made in China.