Editor’s note: With the Trump administration’s attacks on Chinese technology firms, investment in chip suppliers on the Chinese mainland has poured in. A Shenzhen-based startup that’s riding the wave and pushing innovation has developed a chip with a utilization ratio that’s much higher than most chips on the market.
From chess to poker, artificial intelligence has defeated human masters, but the seemingly perfect technology has a hidden handicap — the utilization ratio of AI chipsets averages merely between 10 and 30 percent.
This means that up to 90 percent of the chips’ peak performance can’t be transformed into actual cultivation ability, said Niu Xinyu, founder of Shenzhen-based startup Corerain Technologies.
Niu’s startup has developed an innovative chip with a utilization ratio of up to 95 percent and is in mass production.
The global AI chip market has seen booming revenue in recent years and is forecast to reach US$10.14 billion this year and US$83.25 billion by 2027, according to data from German online statistics portal Statista.
China is one of the world’s largest semiconductor markets, but relies highly on overseas manufacturers, and hundreds of millions of dollars go to imported chips annually.
As the United States moves to curb the growth of Chinese technology enterprises, investment in chip suppliers on the Chinese mainland has seen rapid growth.
Wei Shaojun, vice-president of Tsinghua University’s School of Information Science and Technology, told a forum in July that domestic alternatives ought to replace laggard products with advanced ones, otherwise, it would be a waste of resources and funds.
“If the substitute comes at the same level, it’s better to renovate,” he said. “So, it’s not an easy road to meaningfully supersede.”
He Lizhong, an analyst at Guosen Securities, wrote in a report published on July 20 that a new generation of Chinese chip producers is emerging.
The first generation of chipset suppliers mainly followed foreign bellwethers in terms of hardware, software system and applications. They have the advantage of cost but lack a comprehensive client ecosystem and a strong voice in the international market, He said.
In comparison, he said the second generation of chip suppliers, represented by AI chip startups, has been renovating from the beginning and plays an active role in the ecosystem and industry solutions.
Niu described the chips his company produces as “innovation from the beginning on another path”.
“We rely on technology innovation to provide better performance with lower chip costs,” he said.
Niu explained that most AI chips in the market now adapt the Von Neumann architecture where computation is governed by the instruction execution order, and one computing unit needs to wait for the completion of the last instruction to continue. Its advantage is suitable for universal circumstances, but the method also creates idle time.
Corerain takes another approach — streaming architecture — and launched the world’s first commercial artificial intelligence chip based on such technology in June this year.
The startup’s chief science officer, Wayne Luk, a professor at Imperial College London and a fellow of the Institute of Electrical and Electronics Engineers, is a pioneer in the new AI path globally.
According to Niu, the performance of his accelerator that powers the new Custom Artificial Intelligence Streaming Architecture chip is 3.9 times that of chipmaker heavyweight NVIDIA’s T4 GPU, and its chip utilization ratio can achieve 11.6 times the latter.
Translating the technological efficiency boom into an application example, Niu said if 100 servers based on the Von Neumann architecture are needed to analyze a certain amount of video clips, CAISA only needs about 26 servers.
Compared to Von Neumann architecture, CAISA’s computation order is determined by the data streaming network instead of the instruction execution order, so that the idle time is substantially minimized by cutting the waiting time for instructions.
“It’s just like how the orchestra works,” Niu explained. “Cello players don’t need to wait for the violin to finish, but just start according to the rhythm of the music.”
He said all instrument players coordinate with each other along with the “beat” and Corerain also developed the “beat” in its AI chip architecture, which is measured by nanoseconds, so that data computing and flow can overlap in an orderly fashion.
Niu believes it’s a novel road toward the next generation of AI computing platforms.
Moreover, the startup has achieved mass production in such chips and is conducting tests for autonomous driving, smart cities and intelligent manufacturing.
Leading world IT companies like Intel Corp, Inspur Group and Dell have teamed up with the Corerain for CAISA’s application in future.
