AMT is Developing Neural Network AI Chips for Portable Devices ​

2018-05-31

  China Telecom published Artificial Intelligence Device White Paper on May 17th, 2018.Meanwhile, the first type of AI smartphone has appeared in the market. We are witnessing the dawn of the age of AI. Artificial intelligence has become a major leading strategic technological breakthrough, and it is deeply integrated with the real economy, forming a new kinetic energy to drive economic development, furthermore, accelerate the promotion of humanity into a new era of intelligence.

  Artificial intelligence is to give the machine deep learning ability and the ability to correct the error by itself. At present, it mainly uses computer to simulate the process of brain thinking, thus generates artificial neural network algorithm, so that the machine can think like a human. The most important thing for the machine to learn deeply is data and calculation. Whoever has more data and faster computing, will take the advantage in the competition.

  Current common methods of implementing artificial intelligence, including traditional CPU processor architectures such as X86 and ARM, often require hundreds or even thousands of instructions to complete a single neuron's processing. This pattern seems to be too clumsy for computational requirements of deep learning which do not require too many program instructions but massive data. Especially with the current limit of power consumption, it is impossible to speed up the instruction execution speed by increasing the CPU frequency. This contradiction is increasingly irreconcilable. Deep learning researchers urgently need an alternative hardware to meet the computing needs of massive data. Therefore, the graphics processor GPU has emerged, later the Field Programmable Gate Array (FPGA) is invented. However, all these devices consume much more power, have bigger size, and are less portable. Therefore, currently most of AI operations are performed on the cloud server and then the operation results are sent to the mobile devices. As a result, there emerged the so-called AI processors, yet which are mostly commercial concept hype.

  In order to overcome the shortcomings of the above methods, another type of chips which implements the existing AI capability, that is, the neural network AI chip, has attracted more and more attention in recent years. The neural network AI chip is designed to mimic the human brain. As the human brain's memory is composed of neurons and synapses, it requires only a small amount of energy to learn. Currently, neural unit functions can be implemented with non-volatile memory cells, but it is still uncertain whether synapses can be implemented with semiconductor components. Because the neural network AI chip consumes less power than the existing AI chip with learning ability, it is more suitable for the application on mobile devices.

  Dr. Chung Lam, CTO of Jiangsu Advanced Memory Technology Corporation (AMT), ex-director of PCM project in IBM, is leading four R&D teams from US, Beijing, Hong Kong and Taiwan on neural network AI chips based on PCM technology. The company's product line is as follows: EEPROM/NOR Flash/2D cross point/TCAM/Neural Network AI/3D cross point. It is expected to launch a neural network test chip in 2020 and mass production of neural network AI chips in 2021.

                                                                            By: Ben, Chung Lam, Tess       


本网站由阿里云提供云计算及安全服务