From the official shipment in May 2022, about 10 products based on the WTM2101 chip have been completed by customers. The cooperative customers are mainly various smart wearable device manufacturers with high requirements for computing power, power consumption, and application diversification. Nearly a year after entering the market, Zhicun Technology has shipped nearly one million TWS headsets, VR headsets, smart watches and other customer products. 


As a new type of computing power, the integration of memory and computing can effectively solve the problems of "storage wall" and "power consumption wall" under the traditional von Neumann architecture, and is expected to become one of the advanced productive forces in the era of digital economy.


Recently, a reporter from the "Entering the Industry Chain" column of Chip Myna interviewed Wang Shaodi, founder and CEO of Zhicun Technology, a leading company in the integration of memory computing in China. Discuss the current development of Zhicun Technology companies and their prospects for the future development of the industry under the background of the rapid development of AI chips in the post-Moore era.




The leader in the industrialization of storage and computing, the shipment volume has approached one million in less than a year


It is understood that Zhicun Technology was established in 2017, focusing on the field of memory-computing integrated chips, and is a pioneer and leader in the industrialization of memory-computing integrated.


As early as 2012, the company's founding team began to explore the memory-computing integrated chip technology based on NOR Flash. After 4 years of research in the laboratory, the company's CTO Guo Xinjie went through 6 tapes to capture the floating gate memory-computing integrated deep learning chip technology for the first time, and made the world's first floating gate memory-computing integrated deep learning chip, proving that the memory-computing integrated technology can be used in deep learning.


In 2017, the integrated storage and computing technology began to rise in the industry, and major international semiconductor design, equipment and production companies began to invest in this technology route. Wang Shaodi realized that the integrated storage and computing technology was about to start its era, so he ended his postdoctoral research work early and returned to China to start a business in order to realize the commercialization of the existing integrated computing technology.


After more than 2 years of research and development, Zhicun Technology launched the company's first intelligent voice chip WTM1001 in 2020 based on the memory-computing integrated chip technology of NOR Flash.


WTM1001 is the world's first memory-computing integrated architecture chip product. The chip runs with a power consumption of 300μA and 20 times the computing power of similar chips on the market. It can directly store and run neural networks without data handling during the operation process. Compared with artificial intelligence chips with traditional architectures, the efficiency is nearly 50 times higher. It can be widely used in low-power AIoT scenarios such as wearable devices and smart end point devices.


Wang Shaodi introduced.


The first generation of products allowed Zhicun Technology to run through the mass production process of in-memory computing chips, and achieved a breakthrough from academic research to commercial application.


With the experience of the first generation of products, the company's team started the research and development of the second generation of products non-stop. In January 2022, Zhicun Technology announced the official mass production of the WTM2101, an integrated SoC chip for memory and computing, and successfully launched it to the market in March.


As the world's first SoC chip with integrated memory and computing, WTM2101 is based on the integrated memory and computing technology and adopts a very small package size, which can realize the recognition of hundreds of voice command words at the micro-ampere level, and can realize NN voice noise reduction algorithm with ultra-low power consumption., health monitoring and motion recognition algorithms, the working power consumption in typical application scenarios is at the microwatt level. In addition, the chip can use sub-mW power consumption to complete large-scale deep learning operations, which can cover automatic speech recognition, speech enhancement, health monitoring, environmental recognition, far-field wake-up, motion recognition, visual recognition and event detection.


11.jpg

Zhicun Technology WTM2101 memory and computing integrated SoC chip


Source: Zhicun Technology


The WTM2101 chip has both ultra-low power consumption and high computing power. Compared with the NPU, DSP, and MCU computing platforms, the AI computing power of the WTM2101 is increased by 10-200 times, and the power consumption is only between 5uA and 3mA. Especially for the computing power above Gops, such as the computing power generated when TWS, walkie-talkies, hearing aids, etc. need to use voice enhancement, noise reduction, and anti-howling algorithms, the WTM2101 chip based on in-memory computing technology can give full play to its advantages.


Wang Shaodi added that, in addition, the WTM2101 chip is a general-purpose chip that can be developed. It adopts a 2.6x3.2mm2 extremely small package and includes 1.8M memory and computing integrated NPU, CPU, accelerator group, RAM, etc. Even conventional non-AI algorithms can run with ease. These advantages enable WTM2101 to better meet market demand and quickly realize market promotion.


From the official shipment in May 2022, about 10 products based on the WTM2101 chip have been completed by customers. The cooperative customers are mainly various smart wearable device manufacturers with high requirements for computing power, power consumption and application diversification. It has not been a year since it entered the market, and Zhicun Technology has shipped nearly one million TWS headsets, VR headsets, smart watches and other customer products.


12.jpg

Main product features of Zhicun Technology WTM2101 memory and computing integrated SoC chip


Source: Zhicun Technology


It is worth mentioning that in addition to chips, Zhicun Technology can also provide customers with rich software tools, which can help more developers develop mature solutions based on WTM2101. Not only that, but WTM2101 also provides intelligent voice and health algorithms and continues to upgrade to always maintain industry-leading results.


Following the mass production of the WTM2101 chip, the company has developed the WTM8 series of chips for higher performance scenarios such as object recognition, classification detection, and video enhancement to meet the needs of various energy-efficient and complex edge computing scenarios.


Wang Shaodi said that the WTM8 series covers 2D (1080P-8K) and 3D video processing scenes, and has the core advantages of high computing power, low power consumption, high energy efficiency and low cost. The computing power can reach dozens of Tops, which is relatively high. The second-generation chip product has a 600-fold increase in computing power. According to it, the product has been verified by tape-out and is expected to be put into the market in 2023.




In-memory computing chips have a lot to do on the AI side, and shipments will achieve a leap of tens of millions under the blessing of the domestic complete industrial chain


As we all know in the industry, modern computers generally use the von Neumann architecture, and the memory is separated from the calculator. When performing operations, the original data source needs to be transferred from the memory to the calculator. After the calculator completes the calculation, the resulting data is returned to the memory, which leads to the problems of low computing energy efficiency, slow speed and high power consumption.


As a new computing architecture, the core of memory-computing integration is to fully integrate storage and computing, which effectively overcomes the bottleneck of the von Neumann architecture, and combines advanced packaging, new memory devices and other technologies in the post-Moore era to achieve an order of magnitude improvement in computing energy efficiency. It has become the focus of the current research on memory-computing integration technology.


In terms of end point, according to IDC's forecast, the number of global IoT devices will exceed 40 billion in 2025, and the amount of data generated will be close to 80ZB. In many scenarios such as smart cities, smart homes, and autonomous driving, more than half of the data needs to rely on end point local processing., the AI computing power requirement of a single device is about 0.1~ 64TOPS.


With obvious advantages in power consumption, computing efficiency, etc., the in-memory computing chip can provide higher computing power and lower power consumption per unit area, thereby prolonging the working time of the device, and has broad application prospects on the end side. It can be applied to scenarios with high requirements on runtime, power consumption, and portability, such as home gateways, industrial gateways, cameras, and wearable devices.


If you want to start a business successfully, the judgment of the market is the decisive factor. When the company first started, the big market voice at that time was that AI did not need the end side, only the cloud was enough. But we think there must be an end-side market, because we know what the technology can bring, whether the upstream and downstream supporting things can be realized, and what kind of improvement can be brought to the product after implementation. Therefore, after full consideration and analysis, we decided to cut in from the end-side chip that was the least optimistic in the market from the beginning of the business, and gradually deployed the cloud edge, and this positioning has continued until now.


Wang Shaodi said that in this field, we are the earliest pioneers and leaders in the in-memory computing market, and have formed very strong competitiveness in products and ecology. In the future, we will further improve the universality and stability of the chip architecture, and continuously optimize the synergy of software and hardware to strengthen the company's competitive advantage.


13.jpg


Zhicun Technology Internal Computing Product Planning


Source: Zhicun Technology


As a relatively new chip design method, memory-computing integration also has many difficulties and challenges in the process of moving towards commercial application.


In this regard, Wang Shaodi pointed out: In the process of commercialization from 0 to 1, the memory-computing integrated chip has no mature experience in device research and development and manufacturing, circuit design, chip architecture, EDA tool chain to software algorithm ecology and many other aspects. Reference. In the mass production process, stability, process deviation and other aspects require the team to explore slowly. During this process, we have stepped through many pits and gone through many detours, and finally made breakthroughs in this field. However, these experiences have also helped Zhicun Technology to establish solid and profound technical barriers and first-mover advantages.


In order to speed up the pace of development and further increase its competitive advantage, Zhicun Technology has been very active in the primary semiconductor market in recent years. Just a few days ago, Zhicun Technology had good news again. The company announced that it had successfully obtained a B2 round of financing of 200 million yuan to accelerate the commercialization of memory-computing integrated chips.


Up to now, the company has completed multiple rounds of financing amounting to 800 million yuan. The main investors include SMIC Juyuan, iFLYTEK Venture Capital, Hubble Investment, Paradise Silicon Valley, Shenzhen Venture Capital, SDIC Venture Capital and other well-known institutions.


With the trust and blessing of these well-known institutions, Zhicun Technology will not only achieve mass production of in-memory computing chips in 2022, but also withstand market tests, helping more than ten customers develop and realize functional innovation upgrades based on WTM2101.


Wang Shaodi said that in the future, the company will continue to invest in the R & D iteration of in-memory computing chips and the expansion of market applications, expand shipments and the market simultaneously, and move towards 10-fold growth.


As a leading domestic company that has realized the commercialization of the integration of storage and computing earlier, Zhicun Technology is further enriching the software and tool chain to build a good ecology of storage and computing while iterating around the WTM sequence. It is believed that after the completion of the 0-1 verification period, with the gradual volume of the main products this year and the successive entry of new products into the market, it is expected that the company will enter the fast lane of development.