"I heard about the CPU, the GPU, and the NPU." At the end of November, Hoshin, host Ken Ken, spoke at a scheduled conference in Taiwan at Mate 10, the latest flagship handset for Huawei.
The mobile phone conference held at Dazhi MeiFu Hotel is full of subtle sense of incompatibility.A Ken and fashion singer Guo Shuyao and Xiu JieKai platform, the scene is a tough technical term.
▲ Huawei's flagship machine Mate 10 presentation, from the left repair Jie Kai, Guo Shuyao, host Ken Ken.
For example, NPU (Neural Processor Unit), Chinese can be turned into "neural network processor" is used to deal specifically with the depth of the learning function of the accelerator.
Or, the popular AI chip, the Kirin 970, the world's first AI phone, with a new AI kernel, so the user needs to manually set the shooting scene mode in the past and can now automate it - for example, When you focus on a steak, a knife and fork symbol appears on the screen, indicating that the AI processor recognizes it as food and automatically adjusts to the food's shooting mode.
Apple, Huawei leading AI attack terminal device
The world's third largest mobile phone brand, is showing the current popular semiconductor industry topic "AI on edge" (AI on the terminal device).
Apple's announcement in September this year also revealed that the A11 Bionic chipset on the iPhone X also features an AI-like "nerve-like engine" and, in one fell swoop, "AI on edge."
According to the past as long as the apple is used, it will become the industry "standard" practice, the next high-end handsets of various brands will also be equipped with an AI chip.
A month after Apple's announcement, Counterpoint, a technology market regulator, said in a research report that the market share of smartphones that target AI chips in 2020 will have skyrocketed from 3% in 2017 to 35%. Said that there is one AI chip for every 3 mobile phones.
"The AI is the terminal to come, Huawei and Apple are the first two," also attended the presentation Huawei TSMC IWTO Wang Yaodong, director of senior. And these two chips, using TSMC's most advanced 10 nanometer process.
The concept of the past, will think that a large number of mechanical learning needs computing, need to be done in the cloud, calculate and then return to the phone.
Wang Yaodong said that now the semiconductor process miniaturized to 10 nm, a substantial increase in computing power, enough to be the original cloud computing model shrink to the phone, but for the average general-purpose processor is still overloaded, so add Dedicated NPU. Its computing structure for a simple but a large number of parallel computing, "Wang Yaodong said.
A similar thing happened about 20 years ago when the demand for graphics operations skyrocketed, creating an independent new chip category, the graphics chip.
Huang Renxun, the founder of graphics chip maker Nvidia, first proposed the word GPU in 1999 to simulate the CPU (central processing unit) familiar to the public and let everyone know the importance of the graphics chip.
▲ Huida founder Huaren Founder called the father of GPU, and now he is also the latest AI chip battle grab the initiative, is expected to be listed later next year, the latest AI computing platform Pegasus, claiming to be the world's first to support the fifth Autopilot platform. (Source: Technology News)
Usher in contemporary "Cambrian explosion"
Now in response to AI's demand for the explosion of computer performance, a new chip category has already emerged, but what is it? At present, the industry is incomprehensible, Google called TPU its own server, Huawei called NPU, Intel called NNP (Neural NetworkProcessor).
"This is a big trend," said Zhou Li Dong, vice president of Microsoft Research Asia, which specializes in artificial intelligence systems. "Now it's the XPU era. There are many blooms and lots of solutions come up."
Many people in the industry use the famous Cambrian explosion known as the history of paleontology to describe the lively billowing look now.
Interestingly, the current China's top AI chip design company, called Cambrian technology.The company funds and people team are from the CAS Institute of Computing, by the Chinese media as "AI team" and in August this year Rapid completion of the 100 million US dollars A round of raising funds, shareholders, including Alibaba, Lenovo, although the valuation has not been exposed, but the industry generally believe that the Cambrian has become the world's first AI Unicorn Unicorn field (valued at more than 10 One hundred million U.S. dollars) .
Because, Huawei AI processor Mate 10 core, is from the Cambrian authorization.
"The first option is quicker in the market," said Wang Yaodong. However, he believes Huawei will definitely find ways to use its own AI design next. "Anyone who has the ability to develop his own AI will develop on his own."
▲ Cambrian science and technology originated from the Chinese Academy of Sciences in China, which is Miao Honggen's "AI hope." Perhaps the founding team received very little media coverage.
New breed of processor
The end of October, the global semiconductor industry leaders gathered 30th anniversary of TSMC forum, AI chip has also become a hot topic.At that time, UBS analyst Lu Jia 璈 from the audience asked: "The future prospects of AI chip? Like GPU, DSP become the mainstream IC Category? "
At that time, TSMC chairman Zhang Zhongmou appointed by the most suitable for the question Huang Renxun answer.
Jen-Hsun Huang believes there will be 3 different types of AI chips, the first of which is used to train deep learning models that are computationally powerful and often placed in cloud servers.
The second is the neuro-processor used on Huawei's mobile phones and Apple's mobile phones. "According to Jen-Hsun Huang," a new breed of processors, all future autonomous machines, including unmanned vehicles, unmanned aerial vehicles, People) will have one on top.
Third, the endpoints used in the Internet of Things, close to the sensor, the number of megabytes calculated above.
"In the future there will be AI and mechanics learning on every sensor," added Vincent Roche, chief executive of analog IC maker ADI. "Because a lot of computation needs to be done close to the source of the data."
So many new semiconductor needs, it is also the price of TSMC all the way this year exceeded 200 yuan, the market value of Taiwan's currency soared 6 trillion yuan key.
No wonder, foreign brokerages have recently lowered their TSMC target prices.Zhang Zhongmou can calmly tell the media after attending the breakfast meeting of the AIC. "At present, everyone is talking about AI. AI needs high-speed calculation and development fundamentals Can not see any restrictions ....... (efficient computing) There is no need to slow down the worries. "
Taiwan's newly emerging menace is absent instead?
Currently the most watched in the market, Huida, Intel and many Cambodian-led start-up companies in China, are competing for the UAV, UAV, smart phones and cloud server high Performance AI chip.
Taiwanese manufacturers are still absent from this fertile virgin land.
A senior industry revealed that such high-end chips, the need for R & D team, the use of the process costs are not out .Also go out with Intel, Huida fight .There have been a well-known manufacturer of the team tried to develop, but still in this year Played back, because I think the risk is too great. And even venture capital does not support.
In contrast, China's newly created team, although less skilled and experienced than Taiwan, said, "China now has so much money to raise it," said Lu Chaoqun.
▲ semiconductor veteran, Yuchao Technology chairman Luchaoqun said that in the face of the mainland government's strong investment in new technology of AI, Taiwan's IC design industry still needs the support of all sectors of government.
What's more, the mainland government announced in July that "a new generation of artificial intelligence development plan" will call for achieving the goal of synchronizing AI technologies and applications with the advanced world standards by 2030. It is expected that there will be a new wave of hot money To the AI industry.
"The mainland has invested so much in manpower and material resources that Taiwan has to chase it," Wang Yaodong could not help but say.