ReRAM "memory and computing integration" is a new idea applied to AI large computing power

In the western part of the continent, data centers are rising from the ground, sending a steady stream of computing power to the whole country. But data computing is often accompanied by huge power consumption, so there is a huge contradiction between developing data centers and achieving the goal of "double carbon". At a time when "carbon neutrality" has become a global consensus, the "East Number West Calculation" project launched by the state in recent times has received widespread attention from all walks of life. By laying out 8 computing hubs across the country, "East Counting and West Counting" guides large and super large data centers to gather in the hubs to form data center clusters, so as to optimize resource allocation and improve resource use efficiency. Although the "East, West, And West" provides policy assistance for optimizing the energy supply structure of data centers, it cannot solve the fundamental contradiction between the energy efficiency ratio of existing computing power and the green development needs of computing power energy consumption from the technical level. The power consumption of computing chips has always been one of the most concerned indicators in the evolution of chip manufacturing processes. In the post-Moore era, the path of reducing chip power consumption through the improvement of the process is also becoming increasingly stretched, and it is close to the physical limit of Moore's law. In addition, the huge ecological dependence also involves the development process of various AI application scenarios.

However, with the development and industrialization of ReRAM-based all-digital memory-computing integrated architecture with large computing power, low power consumption and easy-to-deploy chips, this problem seems to have a new solution.

The integration of storage and calculation breaks the "storage and calculation wall" and the "energy consumption wall"

In the process of the development of the chip industry, there are two mountains that cannot be bypassed: "storage and calculation wall" and "energy consumption wall".

In 1946, the Hungarian-American scientist J. Von Neumann proposed the famous principle of storage programs, and since then 70 years, modern computers have developed to the fifth generation, and memory has undergone magnetic storage, optical storage, etc. all the way to semiconductor storage, while von Neumann structure is the most mainstream computer structure the status has never changed. However, with the rapid progress of the times, the inherent defects of the memory-computing separation structure make the von Neumann-structured electronic computer seem to have difficulty meeting the needs of more massive, faster and lower consumption of data processing.

Why is von Neumann's structure inherently flawed? This is because in this structural design, the speed at which the CPU accesses the memory determines the speed at which the system runs, which also makes the system absolutely dependent on the memory. With the rapid development of semiconductor technology, the computing speed of the CPU has far exceeded the memory access speed, the former is usually more than 200 times the latter, and the CPU has to wait for data between executing instructions. The data transfer bandwidth between the CPU and the memory, as well as the manual optimization of each layer of code during the transfer process, have become bottlenecks that limit the further improvement of computer performance, bringing about the problem of "memory wall".

In the process of processing a large amount of data, these data need to be moved frequently in the computing and storage units, and the movement is subject to the technical bottleneck of key components, which also brings a series of problems such as increased chip area, increased interface IP cost, and increased power consumption, which also produces the problem of "energy consumption wall".

Especially when the development of the process began to enter the post-Moore era, the industry became more eager and urgent for a new technical route to solve this problem - so the "integration of memory and calculation" was proposed. This is a non-von Neumann architecture that organically combines storage and computing, and theoretically computation can be carried out directly in memory, which not only breaks the absolute dependence of the system on memory, but also greatly eliminates the overhead caused by data migration, and completely eliminates the problem of "storage wall" and "energy consumption wall".

A new type of storage ReRAM realizes the optimal solution of memory and computing

Since it is theoretically feasible, how can we actually achieve the integration of memory and calculation?

In general, there are about three schools of solutions on the market to achieve memory and computing integration: one is to use advanced packaging technology to package the computing logic chip and memory (such as DRAM) together, to achieve high memory bandwidth and low access cost with high IO density, or to use more advanced process processes, such as 2.5D, 3D packaging and high-level wide HBM interface, etc., to achieve near-memory computing; the other is in the traditional DRAM, SRAM, NOR Flash, NAND Flash implements in-memory computing; the third is to use new storage elements to achieve memory-computing integration.

But unfortunately, most of the technology still can't really solve the problem of "storage wall" and "energy consumption wall".

With the introduction of the "all-digital storage and computing integration" technology developed by ETRUM technology based on ReRAM, the above problems seem to have an understanding method. The material characteristics and production process of the chip are relatively mature, which can break through various problems such as insufficient performance, limited use scenarios, and high manufacturing costs, and it is easier to achieve wide-scale popularization and commercial application.

Not only that, the computing unit based on the memristor (ReRAM) technology can complete massive AI calculations through the storage memory characteristics of the resistive device, using the basic physical laws and principles; through the architecture of memory and computing, it can save the data from the memory unit to the calculation unit layer by layer, so as to save a lot of derivative costs caused by data relocation, which is also in line with the country's technological development route of carbon neutrality and carbon peaking. This has changed the status quo that the amount of data required by traditional AI computing requires huge energy consumption.

In terms of annual growth rate, speed, scalability, compatibility with CMOS, elements of new storage media identified by IEEE, and actual commercialization process, ReRAM is very good compared with other emerging storage technologies such as phase change memory and MRAM.

The strong advantages of ReRAM have attracted a large number of manufacturers to compete for research and application landing, including TSMC, UMC, GROFOUNDRIES, SMIC and Xinyuan Semiconductor, etc. have established commercial ReRAM production lines that can be mass-produced, of which Xinyuan Semiconductor's first domestic 28nm process post-road production line can be mass-produced and commercially shipped.

Yizhu ReRAM all-digital storage and computing integration technology: it has the three advantages of large computing power, low power consumption and easy deployment

In an environment where the ReRAM storage and computing track is flooded with a large number of players, who can get the first-mover advantage of the breakout?

At present, the ReRAM storage and computing track has two solutions, one is analog, one is all-digital, the simulation scheme naturally matches the physical characteristics of the signal, but there is a certain degree of accuracy limitations, in addition, it will also produce the derivative cost of frequent conversion between analog and digital.

As the first company in the industry to adopt an all-digital route to cultivate a reRAM-based memory-computing integrated large-computing chip, E-Casting Technology adopts another: an all-digital solution. Based on ReRAM, the computing power chip of "all-digital memory and computing integration" can improve the accuracy of the product to the greatest extent and solve the problem of low accuracy of analog chips in the previous memory-computing integration technology. In addition, the solution saves the cost of analog technology and results in a solution that guarantees accuracy while optimizing area and power consumption.

The all-digital memory-computing integrated AI computing architecture based on ReRAM can not only effectively solve the problems of memory walls and energy consumption walls, but also solve the "compilation wall" problem faced by current AI applications.

Under the von Neumann architecture, computation relies heavily on multi-level storage architectures, which are very unfriendly to compilers. The compiler can not perceive the dynamic characteristics of the code in the process of compiling the code, there will be stuttering, delay, and the integration of all-digital storage can turn these features into resources directly perceived by the compiler, ensuring the determinism of delay while automatically optimizing the allocation of resources, which can facilitate customers to quickly deploy various types of algorithms.

The latency determination and easy-to-deploy features of the all-digital memory-computing integrated architecture are ideal for autonomous driving scenarios. All kinds of artificial intelligence algorithms contained in automatic driving will have different degrees of delay problems when running, and the technology of storage and computing can not only solve the problem of uncertain delay, but also improve the overall performance of the system.

In addition to solving the delay problem, it also has high density, large capacity and high energy efficiency ratio, which can meet the computing needs of automatic driving and large throughput, and can also meet the requirements of temperature change, vibration and stability of vehicle-grade products in terms of physical stability.

Comprehensive independent research and development of domestic alternatives

From the history of chip development, any widely used good product must work closely with mature processes, such as the joint debugging of TSMC and NVIDIA in history, and constantly optimize TSMC's advanced process process. In the field of all-digital memory and computing integrated AI computing chip of ReRAM architecture, Xinyuan Semiconductor will join hands with Yizhu Technology and closely debug, reproducing the classic case of NVIDIA and TSMC combining swords and jointly developing advanced process processes.

Based on the Metal Wire process, Xinyuan Semiconductor has achieved national production in the design and manufacturing process of ReRAM devices, and Xinyuan Semiconductor has completed the industry's first 28nm process ReRAM chip tape-out, and has built the first pilot line Chinese mainland, with the ability to design and manufacture vertically integrated memory. In addition, Xinyuan Semiconductor uses CMOS-friendly materials, can use standard CMOS processes and equipment, no pollution to the production line, low overall manufacturing costs, can easily make semiconductor foundries have ReRAM manufacturing capabilities, once in the field of automatic driving and AI and other areas of breakthrough applications really started, the process can immediately keep up, which has great advantages for mass production and commercial applications.

In recent years, changes in the international situation have increased the uncertainty faced by the semiconductor industry chain. It also makes localization one of the important construction themes of semiconductor companies.

The combination of E-Casting Technology and Xinyuan Semiconductor will also make the large-computing power and low-power chip-related technologies of E-Casting Technology based on reRAM all-digital memory-computing integrated architecture completely autonomous and controllable, and realize the national production from IP to production.

Billion casting: The small lotus only shows its sharp horns, and there are dragonflies standing on the head.

Although Yizhu Technology is now very young and only began to officially operate in Shanghai in October 2021, it is already the only company in China that independently designs a "memory-computing integration" large-scale computing chip based on a memristor (ReRAM).

At the end of 2021, Yizhu Technology completed the angel round of financing jointly led by Zhongke Chuangxing, Lenovo Star and Huixin Investment (National 5G Innovation Center). It is worth noting that the lead investors are well-known institutions in the field of hard technology investment. So, to be able to play its own advantage in the track of so many giants pouring in, and be favored by so many capitals, what is the strength of Yizhu Technology?

The answer is an international entrepreneurial team composed of senior technical experts, elite engineers, innovation leaders and industrialization pioneers from all walks of life in yizhu:

- The company's senior management team has more than 20 years of experience in research and development, management, marketing and entrepreneurship of different types of chip and system software.

- The company's R & D team members come from many well-known chip companies at home and abroad and many famous universities at home and abroad, such as Stanford, Texas University, Shanghai Jiaotong University, Fudan University and University of Science and Technology of China, etc., and have published 40+ papers at the top of the top meeting with top academic and industrial value, more than the sum of the top papers of other similar track teams in China, and the research results have been commercialized in many well-known head companies.

- The core design team covers all technical fields such as devices, chip architecture, SoC design, software systems and AI algorithms.

- The engineering team also has rich experience in chip design and tape-out, and has excellent capabilities.

Integrating the strong combination of elite talents in the academic and industrial circles, the team of Yizhu Technology has fully covered the physical layer, circuit design, architecture full stack, chip product reference design scheme, commercial landing and ecological construction of the integrated chip.

The AI computing power track in the post-Moore era is expected to win the corner of the new technology

The global demand for AI computing power is growing rapidly. Since 2012, the computing power of AI in the cloud has increased by more than 300,000 times, and it will maintain this growth rate for a long time to come. Because of this, the state vigorously supports the development of related industries. 2020 is the first year of new infrastructure in the true sense, and AI computing power, as a huge driving force for the large-scale development of AI technology and industry, will become the core support of the entire new infrastructure. However, in the post-Moore era, the popularization and application of AI computing power has always faced the constraints of the three mountains of storage and computing wall, energy consumption wall and compilation wall, resulting in the market call for "high computing power, low power consumption, and easy deployment" is also increasing. However, in the objective environment of the first-mover advantage of traditional AI computing power manufacturers, it is difficult for latecomers to catch up on the existing technology track. However, in new technical fields, such as large computing chips based on ReRAM storage and computing, countries are still in the same starting line, and domestic advanced startups such as Yizhu have started.

"The times make heroes." Every big change in computing architecture creates a new king. From IBM in the host era, Intel in the PC era to Qualcomm in the mobile era, the era of intelligent Internet of Things will be the world of contributors to new technology architecture, and we look forward to the new technology of ReRAM storage and computing integration of all-digital computing, which can become a new solution to the storage wall, energy consumption wall, and compilation wall, and win in the corner.