Artificial intelligence is a battle for survival, and the big players are entering a $100 billion market, once again securing victory. The Cambrian has entered the "unicorn" arena, backed by the Chinese Academy of Sciences. The success of a chip depends not only on its efficiency but also on the strength of its ecosystem. Chen Tianshi believes that the Cambrian 1A processor can compete directly with Apple’s offerings.
There's a well-known saying about modern AI: “We tend to overestimate the present and underestimate the future.†If we look back from the perspective of more than 60 years from now, AI has just begun to reshape human society.
From a long-term historical view, and looking back from the future of general AI, today’s pioneers are using the most primitive tools—such as “deep learning†and “neural networksâ€â€”as they explore this vast new territory. Chen Tianshi is no different from others; he walks forward with nothing but his own ideas and effort.
But fortunately, from the stone tools he created, people gradually discovered and preserved the fire.
Cambrian, the startup backed by the Institute of Computing Technology of the Chinese Academy of Sciences, has just entered the market and joined the ranks of “unicorns.â€
Polishing "stone tools"
AI is a revolution. Like the Internet, it will sweep through every corner of life—from nothing to everywhere. Ordinary people don’t even need to chase it; it will come to you, embrace you, and even envelop you. As Luo Zhenyu said, “You don’t have to worry about starting because you will arrive.â€
In this revolution that redefines infrastructure, Chen Tianshi has been an early pioneer. At 16, he was admitted to the Junior Class at the University of Science and Technology of China. In 2014, his paper won the Best Paper Award at a top international conference in the U.S. His company, Cambrian, raised over $1 billion in its Series A round, becoming the world’s first AI chip “unicorn†and the largest AI chip startup globally.
Based on deep learning algorithms, computer vision, speech recognition, natural language processing, and bioinformatics have all seen significant progress. However, traditional processors like CPUs, which are currently the mainstream for AI computing, were not designed with AI in mind, leading to efficiency limitations. Even GPUs suffer from high power consumption and other issues.
To build a neural network as complex as the human brain using a general-purpose processor, you might need to power a whole power station. When AlphaGo was first unveiled, it used 1,000 CPUs and 200 GPUs, costing $300 per minute, and its network size was only one-thousandth of the human brain.
Chen Tianshi believes the ideal AI chip should be a new type of processor capable of handling multiple modalities—like voice, text, images, video, and natural language—with far greater efficiency than CPUs or GPUs. To achieve this, a new set of AI instructions is needed, allowing flexible processing and fast support for various algorithm applications on the chip.
To put it simply, imagine mounting a motorcycle engine (CPU) onto a car (a deep learning platform)—the car doesn’t work. Now, we need to design a specialized engine for the car itself.
The “DianNao†architecture, developed by Chen Tianshi and his team, offers 100 times the performance of mainstream CPU cores while consuming only 1/10 the area and power, achieving a three-order-of-magnitude improvement in efficiency. In 2014, he collaborated with his brother Chen Yunqi and Professor Olivier Temam from Inria to win the Best Paper Award at a top international conference.
Between 2014 and 2016, the team and their international collaborators made groundbreaking contributions to the field of processor architecture, presenting the Diannao series at top conferences—Diannao, DaDiannao, PuDiannao, ShiDiannao, and Cambricon. These innovations achieved hundreds of times the efficiency of current deep learning systems for various tasks.
According to academic analysis, the pioneering work of the “Cambrian†in deep learning processor instruction sets provides critical technical support for China to take a leading position in the intelligent industry ecosystem. Since the introduction of the DianNao architecture in 2014, deep learning processors have become one of the most important research areas at ISCA, with nearly one-sixth of the 2016 ISCA papers referencing Cambrian’s work.
“Currently, software like AlphaGo runs on GPUs. In the future, using the ‘Cambrian’ processor could significantly speed up deep learning (neural network) processes,†Chen Tianshi said.
He described the relationship between general-purpose processors and deep-learning processors using the analogy of a “Swiss army knife†versus a “kitchen knife.†While the Swiss army knife has many functions, the kitchen knife is better suited for cooking. Similarly, in intelligent processing, “Cambrian†is the better tool.
The Cambrian instruction set is specifically designed to handle large-scale neurons and synapses. One set of instructions can process an entire group of neurons and provide specialized support for data transmission between them. Simulation experiments show that deep learning processors using the Cambrian instruction set offer two orders of magnitude better performance than x86-based central processors.
Today, by simulating the computation of neurons and synapses, the Cambrian AI chip intelligently processes information. With a specially designed storage structure and instruction set, it can process 16 billion neurons and over 20,000 synapses per second, while consuming only one-tenth the power. In the future, there may even be a system similar to AlphaGo embedded in a smartphone.
3U Splice Panel,3U Panel,3U Splice Panel Adapter,3U Splice Panel Bracket
Huizhou Fibercan Industrial Co.Ltd , https://www.fibercan-network.com