RCCN WeChat QrCode
Mobile WebSiteAt present, digital technology innovation is inseparable from the development of the real economy. In 2018, China ’s digital economy was 31.3 trillion yuan, accounting for 34.8% of GDP ’s total output value. It is expected that this figure will reach 3.6 billion in 2019. yuan. Behind the rapid development is the technology investment of the Internet and artificial intelligence companies and the application of related industries. A group of excellent artificial intelligence startups are making efforts, such as Shangtang Technology, Defiant Technology, Youbi Technology, Ping An Technology, Ji Chain Technology, etc., AI + Security, AI + Transport, AI + Medical, AI + Video, AI + Finance Combined rapid commercialization. At the same time as the rapid development of the digital economy, the national strategic level has also put forward a plan with more development vision, that is, to provide economic, social innovation, coordination, openness, and shared development. The underlying strategic, network-based basic layout with a multiplier effect "new infrastructure ".
"New infrastructure" is a general term for a series of modern infrastructure systems driven by technology and industry, featuring intensive and efficient, economical, intelligent, green, safe and reliable. The term "new infrastructure" in 2020 is frequently proposed in many important national conferences. Its construction core mainly involves 5G, big data centers, artificial intelligence, industrial Internet, UHV, energy vehicle charging piles, high-speed rail transit In these seven major areas, the breakdown of construction content covers many details such as basic support, platform construction, and industrial applications. Taking artificial intelligence as an example, the main construction directions include: the development of the underlying hardware of the AI chip; the construction of a general intelligent computing platform; the construction of a basic research and development center for intelligent perception processing interaction energy; and the construction of an artificial intelligence innovation development test zone.
There is not much difference in the manufacturing process between AI chips and traditional chips. The difference lies in performance and efficiency. For example, the efficiency of AI chips will be higher when AI algorithms are executed. The efficiency of CPU and GPU algorithms is low, which is still insufficient for commercial use. AI chips serve AI algorithms. In image recognition, video recognition and other fields, CNN convolutional networks are commonly used. In speech recognition, they are mainly RNNs. The front and rear algorithms are very different, but they are essentially matrix or vector multiplication and addition. Then cooperate with some algorithms such as division and exponent.
When the AI algorithm is executed under the same conditions, when comparing IBM's POWER8 (CPU) and Google TPU1 (GPU), the peak calculation times of POWER8 in one second is approximately 64Gops, and the peak calculation times of Google TPU1 in one second is approximately 90Tops.
AI chip will be better than CPUGPU, including light 800, under Rsenet50 v1 condition, peak performance of light 800 computing power reaches 78563IPS; peak energy efficiency reaches 500IPS / w. It can be equivalent to the computing power of 10 GPUs.
At this stage, the combined application of artificial intelligence and other industries is gradually maturing, and the demand for computing power is increasing day by day. In order to match the needs of industrial-level and city-level projects, the demand for AI chips is gradually emerging.
