Just five or six years ago, artificial intelligence existed only in people’s imaginations. Leifeng Network Leifeng Network
In just a few short years, AI has expanded rapidly, and AI computing power and computing data volume are increasing exponentially every year. The demand for computing power is unprecedented, but Moore’s Law is approaching its limit.
Under the obstruction of the energy consumption wall and the storage wall, the increasingly faltering increase in computing power of semiconductors can no longer catch up with the rushing AI.
The integrated storage and computing architecture has the opportunity to solve the problems faced by AI. Under the tide of the times, a group of enterprises exploring the integration of storage and computing have been established one after another.
In this field, domestic and foreign research is almost on the same starting line.
Among them, Yizhu Technology is the “pioneer” that uses the new memory ReRAM for the integrated storage and computing track. Only two years after its establishment, Yizhu Technology has been able to design a ReRAM-based integrated AI computing power chip with an energy efficiency ratio ten times higher than that of mainstream computing cards.
In the upcoming outbreak of storage and computing integration, domestic manufacturers will usher in an upright showdown with foreign giants.
The horse named GPU can’t pull the car of AI
AI has been in development for 70 years since John McCarthy proposed the concept of artificial intelligence at the Dartmouth Conference in 1956.
In the past 70 years, AI has experienced three waves.
In the first two waves, AI ultimately failed to become popular due to various reasons.
It was not until the third wave, the rise of deep learning technology that solved the problems of AI universality and maintainability, that this track really ushered in the dawn, creating a modern sense of “artificial intelligence” based on deep learning. .
With the widespread application of deep learning, the demand for computing power is increasing, which makes GPUs capable of providing large-scale computing power increasingly important in the AI field.
However, GPUs are not designed for artificial intelligence after all. As the development of artificial intelligence enters the deep water area, the problems of GPUs gradually emerge, especially the problems of “storage wall” and “energy consumption wall” in the development of AI, and GPUs cannot break through.
Whether it is a CPU or a GPU, the Von Neumann architecture with separate storage and computing is used.
Under the Von Neumann architecture, 80%-90% of the power consumption occurs in data transmission, and 99% of the time is consumed in the process of memory reading and writing, resulting in the “storage wall” and “power consumption wall” problems.
When the chip of the von Neumann architecture is working, the computing unit must first read the data from the memory, and then store it back into the memory after the calculation is completed, and then the final output can be achieved.
But over the past few decades, the development of memory and processors has been severely unbalanced , and the increase in memory read rates since the 1980s has not kept pace with the increase in processor performance.
This leads to a deformed funnel structure: no matter how much data is processed at the “entry” end of the funnel where the processor is located, it can only be output through the narrow “exit” of the memory, which seriously affects the efficiency of data processing.
Dr. Xiong Dapeng, who used to work in the AI chip company Wave Computing, has a deep understanding of this.
Around 2014, Xiong Dapeng was engaged in research in the field of GPGPU. In his work, he deeply felt that under the constraints of the energy consumption wall and the power consumption wall, it was difficult for artificial intelligence to develop continuously.
Xiong Dapeng believes that a memory-computing integrated architecture that breaks the gap between memory and computing is a possible solution for the future of artificial intelligence.
Different from the von Neumann architecture, the integrated structure of the storage unit and the computing unit eliminates the need for data to be repeatedly “transported” between the two, thus solving the problems of “energy consumption wall” and “power consumption wall”.
Talking about his first impression of the integrated storage and computing technology, Xiong Dapeng said: “I first came into contact with the integrated storage and computing technology in 2017. At that time, I was shocked at how there is such a good thing that can just solve the problems faced by AI chips.”
After he first learned about the integration of storage and computing, Xiong Dapeng was very interested, and immediately began to study and research from the industry to academia, and established a systematic understanding of the integration of storage and computing.
At that time, applying the integration of storage and computing to AI was only a theory, and Xiong Dapeng had not yet found an opportunity to implement his ambitions in the field of integration of storage and computing.
The opportunity for Xiong Dapeng to realize his ambition is the chance encounter with ReRAM, which is also the key to his establishment of Yizhu Technology, an AI chip company that integrates storage and computing in 2020.
The “time of day”, “place advantage” and “people and people” as one
In 2018, Xiong Dapeng came into contact with Crossbar, the leader of ReRAM, at work. At that time, he was struggling to find a solution for the integration of storage and computing technology in the field of AI chips. As soon as he came into contact with ReRAM, he almost intuitively believed that ReRAM had the ability to solve the problem of integration of storage and computing.
There are generally three options for storage media selection: traditional storage media such as Flash; relatively mature volatile memory SRAM; and new memories such as ReRAM.
Xiong Dapeng said that different storage media have different characteristics and their most suitable application fields, and in the scenario of AI large computing power chips that he is optimistic about, ReRAM is the most suitable choice.
In his view, NAND Flash has a large read and write delay, relatively backward performance, and the process node is around 40nm, which makes it difficult to continue to iterate with advanced technology and cannot meet the computing needs of AI high-computing chips.
In the scenario of large computing power, SRAM memory has problems such as limited unit density, leakage current, and difficulty and high cost of engineering implementation.
Although ReRAM also has problems such as accuracy and digital-to-analog conversion when it is applied to the integration of analog storage and computing, in Xiong Dapeng’s view, the fully digital storage and computing integration technology based on ReRAM selected by Yizhu Technology can better solve the accuracy and data. Difficulties such as mode conversion are undoubtedly more suitable for application in the scene of AI large computing power chips.
ReRAM is a new type of non-volatile memory whose basic memory cell is called a memristor, a programmable resistor whose characteristic is that the resistance value can be maintained for a long time after the power is turned off.
The programmable nature of the memristor makes it very suitable for adding computing functions to the ReRAM storage unit, and the feature of keeping data without loss after power failure also makes it a reliable memory, which makes the ReRAM and memory-computing integrated architecture. The requirements coincide.
After a few months with old friends from a global leader in ReRAM new storage technology, as well as well-known scientists from Stanford University, University of Texas at Austin, Shanghai Jiaotong University, Fudan University, University of Science and Technology of China and other universities After the discussion, Xiong Dapeng had a preliminary idea of using ReRAM storage and computing integrated technology to solve the problem of AI large computing power chips, and immediately began to form a team.
Using ReRAM and full digital storage and computing technology to make AI chips is a brand new road. There are few footprints left by predecessors, and there are no stones to touch across the river.
At the beginning, Xiong Dapeng also considered starting from SRAM with more mature technology and then transitioning to ReRAM.
In the end, it was the support of the team that gave Xiong Dapeng confidence. During the formation of the team, Xiong Dapeng found Dr. Debu, who had worked with Wave Computing in the past.
Dr. Debu is an IEEE Fellow and worked at Stanford University. At that time, he was the chief scientist and CTO of the AI department at Cadence. He was also researching the SRAM-based storage and computing integrated IP Core, and has already married. If you accept Xiong Dapeng’s invitation, you will face difficulties in both family and career.
But when Debu learned about Xiong Dapeng’s decision to enter the ReRAM-based all-digital memory-computing chip, he suddenly realized that this was a big future-oriented business that solved the technical bottleneck that he had studied in Cadence based on the SRAM-based memory-computing integration. , to overcome various difficulties and finally decided to join the ranks of Yizhu technology entrepreneurship.
Debu came from a long distance to invest, which made Xiong Dapeng feel the confidence of people of insight in this track, and many experts and scholars from various fields including process devices, circuit design, architecture solutions and software ecology who have had in-depth exchanges with Xiong Dapeng before. The addition of Xiong Dapeng gave Xiong Dapeng the confidence to tackle key technical problems.
“Both ReRAM and MRAM are relatively cutting-edge fields. It is difficult for a company to succeed alone, and it is inseparable from the support of leading partner companies and first-class research teams in the field of new memory.” Xiong Dapeng said.
With these supports, Xiong Dapeng finally made up his mind and decided to start directly with ReRAM.
“Our entrepreneurial team is the ‘three old’ team,” Xiong Dapeng said with a little teasing, “old colleague, old classmate, and old friend.”
Xiong Dapeng, who is very familiar with AI chips, and Debu, who has development experience in SRAM-based storage and computing integration, have made Yizhu Technology start with a favorable location, and the addition of many old friends has added people to Yizhu Technology. And in Xiong Dapeng’s view, the time of the integration of storage and calculation is also present.
The decisive battle is just around the corner, and China Core is facing a “bright sword” this time
For domestic companies with integrated storage and computing chips, the road ahead is still long, but the drumbeat of the decisive battle is quietly approaching.
In addition to being an entrepreneur, Xiong Dapeng is also an investor who has been in the chip field for many years.
As an investor, Xiong Dapeng has witnessed the growth of many emerging technologies. Past experience has told him that the era of the unity of storage and calculation is not far away from now.
In Xiong Dapeng’s view, the integrated storage and computing industry has achieved full scene coverage from small computing power on the end side to large computing power in the cloud, and the supporting facilities of the entire industry chain are maturing.
At the same time, potential customers’ awareness of the integration of storage and computing has become more and more clear, from “not heard of” to “understood” to “expectations” for the product now.
Under the trend of the country’s increasingly strict control of energy consumption, the demand for data centers with high energy efficiency ratio and large computing power is also rising. The ultra-high energy-efficiency characteristics of the integrated storage and computing chip can just meet the market demand.
All kinds of factors are superimposed, and Xiong Dapeng makes a judgment: “From 2024 to 2025, the integrated storage and calculation products will fully bloom.”
There are still three years before the full explosion of the integration of storage and computing. For domestic manufacturers of integrated storage and computing, this is their “time”.
Only two months after the official operation of Yizhu Technology, it received an angel round of financing of more than 100 million yuan jointly led by Lenovo Star, Zhongke Chuangxing and Huixin Investment. Yizhu Technology has also made breakthroughs in the key technologies that promote the implementation of ReRAM.
If the chip is built in an analog or hybrid fashion, the memristor suffers from precision drift and digital-to-analog/analog-to-digital energy consumption as it is affected by the process and environment. This is also a key obstacle to the integration of ReRAM storage and computing.
In order to break through this difficulty, Yizhu Technology chose to tackle all-digital storage and computing integration technology.
The chip is constructed based on an all-digital method, no analog-to-digital and digital-to-analog signal converters are required, and it will not be affected by the signal-to-noise ratio. The accuracy can reach 32bit or even higher. Issues such as IR-DROP.
Based on the all-digital approach, Yizhu Technology will develop the industry’s first ReRAM-based all-digital storage-computing integrated AI high-computing chip. Using chiplet technology, a single module will exceed the 1000TOPS computing power, which is more than four times that of the GPU’s 250TOPS computing power.
Xiong Dapeng said that the product launch of Yizhu Technology is advancing rapidly. The first-generation chip will be launched in 2023, and the second-generation chip will be launched in the same year.
Although everything went well, this is far from the end for both Yizhu Technology and the integrated memory and computing chip.
Xiong Dapeng believes that technically realizing the integration of storage and computing and achieving commercial success are two concepts. In his view, if the integrated storage and computing chip is to be applied on a large scale, it must first establish its own ecology.
Xiong Dapeng told us that in the application scenario of large computing power, the competitiveness of the integration of storage and computing lies in building an ecosystem.
It is far from enough to simply “participate” in the existing ecology. Only by breaking out of the limitations of the traditional architecture and building the overall system with the idea of integrating storage and computing from the beginning can the competitiveness of the integration of storage and computing be truly exerted.
In addition to commercializing the storage-computing integrated architecture in the field of AI large computing power, Yizhu Technology’s goal is to build an ecosystem with other partners on the storage-computing integration track.
In the context of the continuous suppression of my country’s semiconductor field by the United States, the integrated storage and computing chip also carries the mission of breaking through the barriers.
Some time ago, the United States introduced export restrictions on my country’s advanced manufacturing process and high-performance computing design tools EDA, which will undoubtedly bring greater challenges to the future of my country’s AI research.
Xiong Dapeng believes that under the premise that advanced technology cannot be localized in the short term, domestic semiconductors must have the technology to develop decoupling from advanced processes under the same performance conditions.
The integration of storage and computing is an effective way to break through the performance limitations of advanced manufacturing processes: Yizhu Technology, based on the mature 28nm CMOS process and the integrated AI large computing power chip designed by existing domestic industries, can already achieve 7nm CMOS advanced process AI chips More than 10 times the energy efficiency ratio and performance.
Moreover, on the integrated storage and computing track, domestic chip manufacturers are not unilaterally catching up.
In Xiong Dapeng’s view, compared with the traditional track, in the field of integrated storage and computing chips, foreign “giants” are not too far ahead on this newly opened road.
“In general, there is not much gap between domestic and foreign in the field of integration of storage and computing, and in some aspects, we are doing faster and better. The domestic integration of storage and computing is basically start-up companies, and start-up companies can work quickly. They are not afraid of competition from foreign giants, but these giants may not have the determination to embrace revolutionary technologies to revolutionize their own lives, and their efficiency may not be higher than ours.” Xiong Dapeng concluded.
For the final result of this upcoming decisive battle, Xiong Dapeng appeared confident: “On the day when the integration of deposit and calculation fully blossoms, we will definitely be able to defeat them.”
On the storage and computing integration track, the first batch of domestic challengers have already started. It is believed that in the “future battle” that will come in the next few years, domestic chip companies will definitely be able to run out of the “navigator” in the storage and computing field. If you want to discuss the topic of storage-computing integrated chip with the author in more depth, you can add the author WeChat Soldier7887 (indicate your intention). In the next article, we will talk about the SRAM-based storage-computing integrated AI high-computing chip. Welcome practitioners and Author discussion.
This article is reprinted from: https://www.leiphone.com/category/chips/JWn4eGT9Cf5BR0wA.html
This site is for inclusion only, and the copyright belongs to the original author.