Solving the "storage wall" problem: opportunities and challenges in the development of computing in memory

Thanks to the rapid development of the semiconductor industry over the past half century driven by Moore's Law, computing power has been experiencing significant leaps. The integration density of integrated circuits doubles every two years, and then this cycle shortens to 18 months. The performance of microprocessors increases by 1 times every 18 months. However, as the economic cost of silicon chips approaches its physical limit, Moore's Law may become invalid in the future. The processing logic based on the Von Neumann architecture, which is based on the instruction set pattern, requires data to be transmitted back and forth between the processor and memory at runtime, resulting in significant power consumption. Nowadays, the shortcoming of processing massive data, especially irregular massive data, is increasingly evident in high-performance computing environments such as artificial intelligence. Solving the "storage wall" bottleneck has become urgent.In this context, integrating the data storage unit and the computing unit can greatly reduce data transfer, greatly improve computing parallelism, and thus greatly improve the energy efficiency of computing in memory technology, which is being mentioned more frequently. At the same time, industry forecasts also believe that computing in memory is one of the key technologies to solve computer performance and AI computing power bottlenecks.In response to this technology, a number of excellent technology companies have emerged in China in recent years, among which Witmem Technology plays an important role. According to Witmem Technology's founder and CEO Wang Shaodi, Witmem Technology's chief scientist Guo Xinjie started developing computing in memory chips in 2012, and completed the verification of the world's first flash computing in memory chip in 2016, which is also the first computing in memory chip that can perform multi-layer deep learning networks. With such achievements, Witmem Technology was formally established in 2017 and focused on researching computing in memory technology.Regenerate response Different from traditional computing, computing in memory the matrix multiplication computation with the memory unit. It has significant advantages in energy efficiency, speed, computing power, and cost. Since nearly 99% of artificial intelligence relies on matrix multiplication, computing in memory  is a natural fit for AI chips. Based on these advantages, Witmem  Technology is dedicated to developing computing in memory chips.Market demand to a certain extent drives the research and development direction and environment. Various computing demands are emerging in smart scenarios, from wearable products like smart earphones and watches to intelligent cars that require massive computing power. Wang Shaodi understands that computing in memory technology is likely to become an irreplaceable presence in more and more scenarios. He said, "As an emerging technology, the development of computing in memory also needs to follow physical rules. The development of memory over the past few decades has also evolved from small to large capacity, from low speed to high speed, from low reliability to high reliability, and computing in memory technology is similarly applicable. Consumer electronics are easier to implement at the application level, so Witmem Technology's technology is first being implemented in products that are closer to consumers, such as smart earphones. Based on considerations of computing power development and the establishment of new technology reliability standards, Witmem Technology is strategically laying out corresponding products for various computing requirements for long-term development, while forming high-performance and highly reliable products in the supply chain."In 2022, Witmem Technology's self-developed computing in memory SoC chip, WTM2101, officially entered mass production. This is the first computing in memory SoC chip to enter mass production, truly integrating storage and computation, and has high computing power, low power consumption, and high energy efficiency. Recalling the development process, Wang Shaodi explained that since different environments have different requirements for reliability, it was challenging to improve the reliability of technology and products, so more R&D resources were invested. Currently, the chip has been applied to earphones, AR glasses, modules, and watches, and there will be eight more products based on this chip in the future.Collaborating for Industry Development and Doing What Needs to Be DoneMore companies need to participate in and jointly establish the ecosystem for the computing in memory, in order to bring about greater profits and a larger market. Currently, there are domestic enterprises that are deeply engaged in this area of computing in memory technology, but each has different research and development directions. There is already considerable competition among peers in the industry, based on the different requirements of large, medium and small computing power, and the market also has differences. Wang Shaodi acknowledges that each company has different positioning and goals, and invests resources accordingly. In thecomputing in memory  space, the demand for large, medium and small computing power is also different, and the market also has differences. Witmem Technology will not define itself too early, but rather strive to make its products cover more high-performance markets. At the level of mass production, stability and reliability at each stage need to be considered, and improving each generation of products is what Witmem Technology needs to do at present.Computing in memory technology originated in academia and is applied in various industries. AI computing has also implemented landing technology on a large scale in the industrial sector. Wang Shaodi said that every emerging technology is basically born in academia. After time and in-depth research, through technology transfer by the industry, its capabilities are expanded again. For example, deep learning technology such as AI has tremendous development prospects after being applied in the industrial field. Since 2010,computing in memory  technology has also made great leaps forward due to the joint development of academia and industry. Therefore, Witmem Technology has cooperated with several top domestic universities and research institutes to find better solutions to more advanced scientific problems, jointly promoting the cutting-edge development of computing in memory technology, and working to transform industrial needs into technical indicators for academic research, acting as a bridge between industry and academic research.Sustained investment and the development of cutting-edge technology will inevitably require a large number of excellent talents. Regarding talent reserves, Wang Shaodi has his own insights. He said, "In different fields, Witmem Technology continues to search for suitable technical R&D talents, and will also provide continuous training for relevant technical personnel within the company. At the same time, we are building a more diverse team of experts, so that each field has its own 'expert-type' talents."In the post-Moore's Law era, more and more new technologies are emerging in the industry. Every technology requires a large amount of verification and continuous innovation and iteration in order to bring about truly quantifiable landing products. At the same time, every technology faces enormous challenges. Not only is the technology itself difficult, but the matching solutions and large-scale commercial opportunities for the upstream and downstream of the industrial chain also face challenges. The establishment of a complete ecosystem and platform, as well as the value of the technology itself, will also affect the development direction of the entire industry. This is also the reason why Wang Shaodi never stops moving forward.