TextChapter from WeChat public number: CV intellectual (ID: CVAI2019) , author: Han Jingxian, editor: Zhang Lijuan, cover from the visual China
At present, both startups and traditional chip makers are under increasing pressure. In one sentence, the AI chip market in 2019 is that the industry tends to be calm.
In 2016, the first generation of AI chips began to break out, and traditional chip makers, algorithm companies, and Internet giants rushed in; now, after three years, “commercial landing” entered the redemption period.
“You can still talk about AI chips in the last two years, but if you don’t have a product made this year, you are inferior to your competitors.” Investors from the semiconductor industry Yang Lei, managing director of the company, expressed his cognition to CV.
But the process of “landing” is obviously more challenging than the chip development itself. This is both a touchstone for the first generation architecture design and a huge force for software development and customer support.
Of course, some AI chip companies have just launched or just started to land products and landing scenes, but whether they can really be applied to products in batches, whether they can really meet the actual scene requirements, and the stability of the chip remains to be seen.
But this does not prevent market research organizations from continuing to portray the bright future of AI chips: According to Gartner’s forecast, the market size of global AI chips will increase from $4.27 billion in 2018 to $32.3 billion in 2023, 2019. The average growth rate in 2023 is about 50%.
No one doubts the future of artificial intelligence, and no one doubts the industrial roots of AI chips, justOn this battlefield, which has its own objective laws, fast and slow, how do participants adjust their posture battles? The first-in-comers have a lot of crowns. Does the latecomer still have a chance? CV Intelligence chatted with many practitioners in the industry with curiosity.
landing
Objectively, in 2019, the pace of commercial use of AI chip players was not as fast as expected.
Whether it is huge, such as Huawei, Ali, Tesla, or those fast-running startups such as Yizhi, Denglin, Sugawara, Jinyun, etc., even if some have already released products or announced the release of the film, but There is still some distance from mass production, not to mention the majority of the giants’ products are still used by themselves.
There are people in the industry who even ridicule. A year ago, at a summit, ten startup companies showed their AI chip plans with PPT. This year, these companies just showed the updated PPT.
The commercialization problem faced by AI chips can also be seen from the current financing case. One investor said to CV Intelligence that several mainstream AI chip start-ups that are currently merging money have a commonality behind them: Endorsed by customers or industry resources, “otherwise it is difficult.”
In the view of Niu Yuyu, founder and CEO of Shenzhen Jinyun Information Technology Co., Ltd., “It is not so difficult to commercialize AI chips, but more than enough to understand the difficulty of landing AI chips.”
The core product specifications of the AI chip are only two: the cost performance and algorithm support versatility.
These two metrics support the continual reduction of AI application landing thresholds, including deployment costs and development cycles. Around these two core indicators, there are subdivided indicators such as domestically replaceable, domain-specific interfaces, and field heat dissipation power stability indicators.
“The process of landing the AI chip is actually the process of releasing the difference between these indicators and the customer’s existing product specifications to the customer.”
This process has its own objective laws and no shortcuts. “There are systematic and engineering pits for product mass production, product introduction, and product shipment.”
It’s like making a car’s AI chip. “The car itself is a safe thing, so the car’s AI chip must pass the car-level inspection, and the system function and safety must be fully verified and cannot be crossed.” Seiko Intelligent Technology co-founder and COO Liu Weihong expressed his cognition to CV.
Yang Lei uses the “hurdle” to describe the AI chip startup. “The hurdle consists of three elements: the athlete’s own ability (the entire team’s configuration) , the height of the column (difficulty of the AI chip) and the length of the run-up (objective Development time).”
Of course, when the outside world questioned the slow landing of AI chips, some industry insiders said that whether it is from an emerging field such as AI, from an underlying hardware platform, or from a chip, it is a good business to have a landing product. Progress has been made.
Cloud: Fairy Fighting
According to the scene, AI chips can be divided into cloud and terminal.
The cloud AI chip has strong performance and can support a large number of operations at the same time. In addition, it can support a variety of different applications such as pictures and voice.
A recent release by market research firm ABI