What is the progress of domestic NPU chips?

fish001

What is the progress of domestic NPU chips? [Copy link]

 

In recent years, the research of NPU by chip companies such as Huawei and Cambrian has become the focus of attention in the industry. NPU, also known as embedded neural network processor, adopts the architecture of "data-driven parallel computing" and is particularly good at processing massive multimedia data such as videos and images.

Advantages of NPU chips

NPU is a neural network processor that simulates human neurons and synapses at the circuit level, and uses deep learning instruction sets to directly process large-scale neurons and synapses. One instruction completes the processing of a group of neurons. Compared with the von Neumann structure of CPU and GPU, NPU integrates storage and computing through synaptic weights, thereby improving operating efficiency. However, NPU also has its own defects, such as not supporting the training of a large number of samples.

Many top AI technologies are criticized for their high power consumption. IBM's "Deep Blue" in the 20th century and Google's AlphaGo in 2016 need to be supported by huge data calculations. The former uses supercomputers and the latter uses server clusters, which cannot be separated from the computer room with constant temperature and humidity. AlphaGo costs $3,000 in electricity for a single game of chess. Zhang Yundong called them "a scientific experiment" that is still a long way from the landing and application of technology. This highlights the advantages of miniaturization, low power consumption and low cost of embedded NPU, and accelerates the landing and application of AI technology. For example, drones have high requirements for the weight and power consumption of cameras, otherwise it will affect take-off and endurance.

Domestic NPU chip process

On June 20, 2016, the National Key Laboratory of Digital Multimedia Chip Technology of Vimicro announced in Beijing that it had successfully developed China's first embedded neural network processor chip, becoming the world's first embedded video acquisition compression encoding system-level chip with deep learning artificial intelligence, and named it "Starlight Intelligent No. 1". This deep learning-based chip is used in face recognition and can achieve an accuracy rate of up to 98%, exceeding the recognition rate of the human eye.

In 2018, Alibaba DAMO Academy developed a neural network chip, Ali-NPU, which will be used for AI reasoning calculations such as image and video analysis and machine learning. The price-performance ratio of this chip will be 40 times that of similar products. In the future, AI intelligence will be better used in business scenarios, improving computing efficiency and reducing costs. CPU and GPU are general-purpose computing chips designed to process thread logic and graphics. They consume high power and have low price-performance ratio when processing AI computing problems. In the field of AI computing, dedicated architecture chips are urgently needed to solve the above problems.

On October 31, 2018, Hangzhou Guoxin took the lead in releasing the IoT artificial intelligence chip GX8010 equipped with NPU, which attracted widespread attention in the industry. As early as the beginning of 2016, Guoxin people began to develop neural network processors and invested heavily in in-depth research. They not only completed the first generation of neural network processor gxNPU, but also completed the design and mass production of the overall SOC chip.

Cambricon has released the high-performance machine learning processor chips "Cambricon MLU100" and "Cambricon MLU200". The two chips are mainly aimed at the intelligent processing needs of the server side, for inference and training respectively. Different from the common name of "neural network processor", the new server chip product is named "machine learning processor".

在华为全联接大会2018上，华为发布了两款AI芯片和全栈全场景AI解决方案，正式打响了进攻人工智能的号角。去年在德国柏林的IFA展上，华为正式发布了麒麟970芯片，该芯片中首次内置了神经元网络单元以完成人工智能计算。同时发布两款AI芯片，华为昇腾910和昇腾310，均采用华为自研的达芬奇AI架构，属于全球第一个覆盖全场景的人工智能IP和芯片系列。

qzcp

China still needs to work hard

alan000345

This is really exciting good news.

zxopenljx

Thanks for sharing