Why are AI chips “cold and hot”?
Is the AI boom over? The judgment that many people who pay attention to AI may give. There is also some basis for this judgment. AI chips, one of the three major elements of AI development, appeared in a large number of startups around 2016, and then gradually cooled down after a hot period of about two years.
However, starting in the second half of 2020, there has been more financing news for AI chip startups. During this month’s WAIC 2021, many cloud AI chip and edge AI chip startups have released new products, and AI chips seem to be hot again. stand up.
Why is the AI chip “cool” for a few years and then “hot”? What is the key to the popularization of AI? Will RISC-V chips become the protagonists of the AIoT market?
Poor experience AI products hurt consumers
There are two key events in the new round of AI boom around 2017. One is that AlexNet won the image recognition competition with convolutional neural network (CNN) in the ImageNet large-scale visual recognition challenge in 2012, which once again set off a wave of AI research in the academic community. In 2017, AlphaGo defeated Ke Jie, the top human player, and swept the entire Go world. The upsurge of AI has spread from academia to industry, and has aroused the attention of the whole people.
“I think AI can become popular again because neural networks can solve some problems after technological breakthroughs, and this can just satisfy consumers’ yearning for a smart and better life.” Founder, Chairman and Chief Executive Officer of Canaan Technology Executive Officer Zhang Nangeng said.
Soon, many AI products such as AI speakers, AI Robots, AI home appliances, and AI story machines were quickly introduced to the market. Among them, the AI speakers with the largest shipments quickly became popular, but were quickly abandoned by consumers.
“Reducing costs by reducing materials and launching products with poor experience is very harmful to the market. Because there is no good experience, it is difficult for the scale of the entire market to grow. Cost reduction should rely on better technology.” Zhang Nangeng said, “AI products It still didn’t meet the expectations of consumers, consumers want a complete product, which is why I think AI has cooled down again.”
Song Jiqiang, vice president of Intel Research Institute and dean of Intel China Research Institute, also told us: “Many AI systems now have only one prototype, but when commercialized, the prototype will not work. For AI innovation to truly develop on a large scale, it must be Vertical integration has to be done very seriously.”
That said, there is still a gap between the actual experience of AI products and what consumers expect. From the perspective of AI intelligent hardware companies, with the support of AI chips and AI algorithms, they have launched products that are more intelligent than ever, and their product performance has been significantly improved. However, from the perspective of consumers, AI products are still not smart enough, and even a little stupid, and it is difficult to continue to use or buy new AI products.
“Another issue is price. For smart hardware and white goods, consumers’ requirements are world-class performance, but price expectations are ‘hell-level’. This is the contradiction.” Zhang Nangeng said. “However, as long as there are clear needs and goals to achieve world-class performance, while taking into account an acceptable price is not an unattainable goal, Canaan hopes to achieve this goal in the next 1-3 years.”
It should also be noted that edge reasoning is the key to the large-scale implementation of AI applications. This is because cloud-edge collaboration has become a mainstream computing paradigm. In the cloud-edge collaboration architecture, edge AI can effectively make up for the shortcomings of latency and data privacy in data desensitization, data perception, and real-time decision-making, and can also reduce high IT infrastructure costs.
At this point, higher-performance AI chips are critical.
Demand is released, and the absolute value of the edge AI chip market increases rapidly
Canaan has a first-mover advantage in the AI edge chip market. In 2018, it launched the first AI inference chip, Kanzhi K210. This AI SoC did not have MobileNetV2 and YOLOv3 algorithms when it was launched, so it only had 1TOPS of AI computing power. “Due to the early research and development time of Kanzhi K210, the algorithm model that appeared later was not considered in the computing power planning, resulting in the limited expansion of this chip in application scenarios. “Zhang Nangeng commented on this chip.
In the context of the entire industry exploring AI chips and AI applications, Canaan’s first AI chip has won orders from multiple head education robot customers. Considerable revenue has also been achieved in the context of the global epidemic and lack of cores in 2020.
“This year, the market demand for many applications has increased rapidly. In the first half of the year alone, we achieved more than 13 million AI chip revenue. If the production capacity problem in the second half of the year cannot be solved, I expect the revenue to be similar to the first half of the year.” Zhang Nangeng said, “I see What’s more, the market demand is particularly strong this year, and the growth is relatively fast.”
Of course, the AI edge chip market is still growing rapidly in absolute value, but the overall scale is still small.In the fast-growing market, in addition to AI education and the addition of AI-enabled network cameras (IPC), there are more and more innovative products that use edge AI chips for crowd gathering detection, ultrasonic positioning, robot dogs, and follow-up PTZs.
“Kanzhi focuses on AI performance. Based on the capabilities of the K210, there are some innovative applications that we did not expect. Smarter products also require higher-performance AI chips. In fact, they are like sweeping robots in home scenarios. The environment is also very complex, requiring relatively high computing power and cost control.” Zhang Nangeng said.
To this end, Canaan launched the Kanzhi K510 edge inference chip. The CPU still uses RISC-V (dual-core 64-bit). The new KPU improves the AI performance to 3TOPS, supports TensorFlow, PyTorch and ONNX model import, and cooperates with KPU to accelerate AI. DSP has also been added.
In addition, Kanzhi K510 supports three-way sensor input 1080P video, and has many high-speed and low-speed interfaces. Among them, the high-speed interface DDR PHY and MIPI RX/TX DPHY are IPs with completely independent intellectual property rights of Canaan.
From the time of launch, Kanzhi K510 is more than half a year later than Canaan’s product roadmap.According to Leifeng.com, the reasons include that this is Canaan’s first relatively large-scale AI chip, the AI acceleration unit KPU has been almost redesigned, the high-speed interface is also completely self-developed, and multiple factors such as shortage of production capacity have led to the investigation of The Zhi K510 was launched a little later than expected.
“It will definitely be better to launch the product earlier, but it is also beneficial to be a little later. When we launched the first AI chip, most small and medium-sized customers could only use the original chip for some embedded development. This year, we have the ability to There are many more customers developing AI chips, and many customers have their own AI algorithms. The market is maturing and application scenarios are increasing, which is a good thing for the promotion of our products.” Zhang Nangeng explained.
Unlike traditional functional chips, developers of AI chips do not know their specific needs, which makes it more difficult to define products for AI chip companies that want to predict market demand in the next three to five years. However, it is clear that as the complexity of AI applications increases, the demand for AI computing power also increases, and improving AI performance becomes the key.
“The KPU of Kanzhi K510 is basically redesigned compared to the previous generation. The original KPU is relatively simple. If the first-generation architecture is used for proportional amplification to improve performance, there will be many problems. The newly designed second-generation architecture is more general and scalable. The performance is better, it is easy to achieve a computing power of a few tenths of TOPS to a few TOPS, and the compatibility is also better.” Zhang Nangeng said.
While improving AI performance, the visual performance of Kanzhi K510 has also been significantly improved. It integrates the latest generation image processing unit and 3D ISP module, supports depth camera TOF sensor, and can automatically parse and extract depth images without software involvement. Accelerates 3D image processing.
Zhang Nangeng pointed out that Kanzhi K510 is a mid-to-high-end computing power chip, and the visual hardware configuration is also the highlight of this chip. It supports input from up to 3 cameras at the same time, one of which supports 3D function, which also caters to market needs, such as living body detection. and so on. The Kanzhi K510 can also be used in high-definition aerial cameras, video conferencing, robotics, STEAM education, Industrial cameras, and assisted driving.
Lei Feng.com noticed that Kanzhi K510 starts from multiple perspectives in reducing power consumption, such as solving the problem of high power consumption and large area of the chip caused by the huge clock tree through the NoC bus architecture; KPU according to different levels of bandwidth, computing and storage requirements It also improves the data multiplexing rate and reduces the power consumption of the chip; it also supports the voice VAD function to meet the needs of low-power design.
However, the key to bringing Canaan’s AI chips to the wider market is RISC-V.
Rush to the trillion-dollar AIoT market
“Ease of use is very important, and developers can implement functions if they use a CPU or DSP, which can reduce development costs and accelerate time to market.“Zhang Nangeng said, “This is also an important reason for us to add DSP on the basis of dual-core 64-bit RISC-V CPU. “
Stronger CPU performance is also increasingly important for a SoC.“There is a very interesting phenomenon. Five years ago, many people in the industry believed that the role of the CPU in the SoC would be weakened, and most tasks would be processed by hardware accelerators, including AI applications. In the past two years, whether it is a mobile phone or a PC , CPU competition is becoming more and more fierce, most tasks are still completed on the CPU, and the CPU has become very important, so the performance requirements of the CPU have also increased significantly.”Zhang Nangeng said.
“I think that in the next few years, the advanced technology and architecture of the desktop and mobile terminals will be rapidly devolved to the original MCU, and this change may not be seen until the next 3-5 years.” He made a prediction.
The importance of the CPU in the SoC has become important again, meaning that the performance and ease of use of the CPU become the key to the competition. In the AIoT market, although Arm, which dominates the mobile market, has a dominant position, the RISC-V instruction set chip has made rapid progress in this market in the past two years. In applications that do not require high software support, RISC-V has been widely used. But applications that require higher performance and more software support face challenges.
In fact, the CPU of Kanzhi K510 is self-developed by Canaan.“We use the RISC-V open source core for the first generation of products. The obvious advantage is that there is no need to buy copyrights to save costs, but there are also many loopholes. The reason why the new generation of Kanzhi AI chips is self-developed is mainly because of our The demand exceeds the demand for open source kernels.” Zhang Nangeng explained why he started to develop RISC-V CPUs by himself.
Self-research of hardware is relatively easy for powerful chip companies, but the construction of software ecology requires the joint efforts of the entire industry. “Whether it is Ali’s work on RISC-V, or if Intel acquires SiFive, it will have a very positive effect on the construction of the entire RISC-V ecosystem.” Zhang Nangeng said: “We are firm supporters of RISC-V, so Whether it is hardware or software, we benefit from the RISC-V community, and of course we also upload our own designs and documents to the RISC-V community, hoping to lead by example to promote the development of the RISC-V ecosystem.”
“The support of more systems has a very obvious effect on the large-scale promotion of our products. When the Kanzhi K210 was first launched, it was very difficult for developers to develop without system support. Later, with the support of RTOS and Linux, this It is very useful for the application of our Kanzhi K510.” Zhang Nangeng said.
“At present, it seems that RISC-V chips may be worse when used in the game field, but for other applications, using RISC-V chips is not a problem.” Zhang Nangeng believes, “But I don’t think RISC-V will be completely Eliminate Arm, and Arm will still have a place in many fields.”
It is undeniable that the ecological construction of RISC-V is still catching up with the ecology of Arm, far from catching up.As far as Canaan is concerned, after they designed higher-performance AI chips, the implementation of many functions was hindered by the lack of software personnel.
Zhang Nangeng said that he has long been aware of the problem of lack of software personnel, but he does not intend to solve this problem by recruiting a large number of software talents. Instead, he maintains the genes of Canaan hardware and solves software problems by working with more professional partners. challenge.
Zhang Nangeng revealed that Canaan will announce the cooperation progress in software as early as the second half of this year.
There is a deeper problem,Why did Canaan choose an immature RISC-V CPU to launch an AI chip? Zhang Nangeng’s answer is: “In the choice of industry, we chose AI, a market with huge demand. In terms of technology, choosing RISC-V has a cost advantage, and it will make a big difference in the future AI market. We’ve always embraced it.”
McKinsey predicts that the global AIoT market will reach $11.2 trillion in 2025.Zhang Nangeng continues to be optimistic about Canaan’s AI revenue of 100 million yuan in the next 3-5 years, and predicts that when Canaan’s AI business revenue reaches a certain point of 5 million to 150 million US dollars, it will usher in a breaking point. .
Why do AI and AI chips cool down after a few years of hotness? It can be said that the landing progress is slow, and further, there is a lack of complete AI products to provide consumers with a good experience. As the key to improving product experience, the improvement of chip computing power is very important. As the first company to launch AI chips, Canaan is also the first company to choose RISC-V. The evolution of its products and technologies provides us with a lot of reference information for the development of the AIoT market.
From using the public version of the CPU to choosing self-developed, from the upgrade of the KPU architecture, to the self-developed high-speed interface, we can see that this market is growing rapidly, and we can also see Canaan’s confidence in the AI market and doing a bigger thing .
Zhang Nangeng said: “The core of a company’s success is the correct strategy. For Canaan, insisting on independent research and development is a strategy to solve things that others have not solved.”
Chip shipments have exceeded 100 million, and Canaan, which already has a 300-person chip team, maintains iterating an AI chip every two years. According to the level of computing power, it is divided into 2 series, 5 series, and 8 series to meet the needs of terminal and edge markets. What surprises will the Kanzhi chip optimized for the scene bring us?
Pre: The country’s first “face... Next: NCIC’s embedded computer contin...