【Text/Observer Network Lv Dong】
In recent years, new workloads such as artificial intelligence, Internet of Things, big data, and autonomous driving have continued to emerge, accelerating the era of diversified computing. In addition to classical computing, cutting-edge computing fields such as quantum computing and neuromimicry computing continue to attract attention.
For the current great power competition, cutting-edge technologies such as quantum computing have become a new competitive highland. For enterprises and scientific research institutions, the key point to measure the ultimate success of a cutting-edge technology is whether it can be industrialized.
In view of the current development status of new computing technology, Observer.com recently interviewed Song Jiqiang, vice president of Intel Research and president of Intel China Research Institute, in which he explained the industrialization degree of cutting-edge technologies such as neuromorphic computing, as well as Intel's special path for developing quantum computing technology, and explained what problems remain to be solved in the industrialization of these cutting-edge computing technologies.
Song Jiqiang believes that the path to the commercialization of quantum computing is not so fast, because there are still many actual "pits" that have not been filled, and everyone is still doing qubits, and the quality needs to be improved. "First of all, the stability of qubits, their fault tolerance, and the length of their coherence time, including how many qubits can be really stable together, these are all questions."
As a traditional semiconductor giant, Intel is not only promoting classical computing, but also laying out cutting-edge computing fields. For example, the company launched a new quantum chip called Tunnel Falls, which contains 12 silicon spin qubits. In the field of neuromorphic computing, Intel has also launched the second version of the Loihi chip, and cooperated with domestic universities and institutions such as Peking University, Fudan, Chinese Academy of Sciences, and Pengcheng Laboratory.
The following is a transcript of the interview:
Observer Network: What is the degree of industrialization of cutting-edge technologies such as neuromimicry computing and quantum computing, and what are the challenges of large-scale industrialization?
Song Jiqiang: From these two points, the path of their industrialization, including time, is still different. In terms of quality or software and hardware maturity, neuromimicry computing is actually closer to industrialization. Because it itself is made by using mature processes, Intel 4, Intel 3, Intel 20A processes can be used in the future. It mainly designs a new set of hardware frameworks for neuromorphic computing, the hardware architecture design is new, and the hardware programming is also new, because this must correspond to the hardware, including the compiler. Therefore, its commercialization difficulty mainly lies in finding a particularly suitable application for large-scale use.
The advantages of neuromorphic computing are obvious. First of all, it can complete some intelligent actions like the human brain, under the condition of saving power consumption. Second, it can do a variety of different tasks at the same time, rather than saying that a piece of hardware can only do one thing, in the future we also hope to be able to support a variety of different modalities on the same chip architecture, a development environment, to support a variety of different modalities to interact with each other, just like the human brain, there is olfactory processing, auditory processing, taste processing, visual processing, cross-correlation with each other.
If it is pure deep learning, it cannot be related, and its model is to do vision is to do vision, and to do hearing is to do hearing. Now, of course, there are also multimodal inputs, but the path is different.
In neuromimicry computing, we actually hope to really simulate the human brain, slowly form different brain regions, these brain regions, when they really constitute some cognitive abilities, can affect each other, for example, you can also know whether this is an apple or durian with your eyes closed, and you can also know through touch and other perceptions, which is the effect that neuromorphic computing wants to achieve.
It does have to go this way in terms of ability, but it must find a breakthrough in the field of application, which application will make it unique, so that everyone can run this application, just like deep learning, it has been waiting for many years, and finally found a suitable time point, from the cost performance, ability, application and other aspects of hardware, in the field of computer vision first run. From the perspective of the relative maturity of neuromorphic computing hardware, Intel is now iterating to the second version of Loihi 2, which is much better than the first version in terms of energy efficiency ratio and programmability. Of course, we have to continue to iterate, and we need one or two key applications to take it flying.
But quantum computing is different, quantum computing is still facing many difficulties, because everyone is still doing qubits, and the quality needs to be improved.
First of all, the stability of qubits, their fault tolerance, and the length of their coherence time, including how many qubits can be truly stable together, are all questions. Once a certain number of logical qubits are formed and can be well controlled, the above algorithms and applications will develop relatively quickly. Now in fact, at the bottom layer, from what way qubits are prepared, to how to solve error control, to how to extend its coherence time, and the future to really do industrialization, what kind of scheme to use to support industrial production, and even what the future quantum computer looks like, these are still inconclusive.
Intel neuromorphic computing chip Loihi
Observer Network: Can quantum computing be said to be in the early research stage?
Song Jiqiang: It's still early days, and if you want to be mass-produced and achieve sufficient fault tolerance, you haven't reached that step. However, in terms of the preparation of qubits, Intel's path is more unique, that is, using silicon electron spin to prepare qubits, which is very different from the industry's long-term preparation of qubits through superconducting solid-state circuits.
Observer.com: In many people's impressions, Intel is a traditional semiconductor giant, but now Intel Research is doing a lot of cutting-edge research, such as neuromorphic computing and quantum computing. What are Intel's considerations when making these cutting-edge layouts? How much power or resources are invested in it now?
Song Jiqiang: What we want to do is technology that can be used on a large scale in the future. Therefore, whether it is to make a quantum computing chip or a neuromorphic computing chip, the final measure of its success depends on whether it can form a large-scale and industrialized effect.
On the road of qubits, why we chose the method based on silicon electron spin is also because we found that this is most suitable for Intel, and industrialized it on the basis of 300 mm, large-scale CMOS preparation of this semiconductor technology. At the same time, we also see that qubits are made in this way, and its size is miniaturized, 1 million times smaller than what is now prepared based on superconducting qubits. Because it is made by CMOS process on silicon wafers, the size of a qubit is not much different from the size of a transistor, so it has the basis for large-scale production. Of course, there are still some things missing, such as how to improve the yield of mass production of qubits, including how to form a computing system in the end. But since we already know that most of the processes are actually similar to semiconductor processes, they are more industrialized.
Neuromorphic computing is actually the same, when we do it, we also have to consider not only like academia, send a bunch of papers, so that doctoral students can graduate, we have to see how to commercialize it in the future. So after we came out with the first test chip, Loihi, we began to build a global community called the Intel Neuromorphic Research Community (INRC). This community is to start with start-ups and scientific research teams that have been doing research in this area in the early days, do test development on test chips, see how the effect is, and then improve the hardware design, so the second version of Loihi 2 performs much better.
We are also working with start-ups and some of the world's top 500 companies to explore the application of neuromorphic computing chips, because when a new chip comes out, it must find the right use. We tried to try different kinds of applications on terminal equipment, on robots, on drones, and on the optimization of cloud data centers, hoping to find scenarios suitable for large-scale use. This is where the industrialization path is different from the academic path. This project, the current research institute basically maintains a team of fifty or sixty people doing it, and there are actually more people doing quantum computing than this, because quantum computing is done by the research institute and the TD (technology development) department at the same time.
Observer.com: Is neuromorphic computing being studied by Intel Research China and headquarters?
Song Jiqiang: The design of the chip is mainly done by the headquarters. Our side is mainly to help them build the Intel neuromorphic research community in China, which includes academic institutions and some enterprises, and we provide them with hardware for them to do experiments, such as the Institute of Automation of the Chinese Academy of Sciences, Peking University, Fudan University, etc., and we also have some experimental applications.
Observer Network: As mentioned earlier, quantum computing is also an area that Intel Research attaches more importance to, what is Intel's position in the entire industry at present?
Song Jiqiang: Intel technology is in the first echelon. In the quantum computing chip, Intel really made it, rather than using a supercomputing system to simulate quantum computing. In addition, we also realize that the path to the commercialization of quantum computing is not so fast, because there are still many practical "pitfalls" that have not been filled. In fact, Intel is very serious about quantum computing, and we are exploring the preparation of quantum computing chips, including how to really make it in a product line with high yield and high mass production capacity.
With quantum chips, we also need low-temperature wireless microwave chips to control it, which Intel itself is doing. It is equivalent to saying that the first must have a computing chip, and the second must have a microwave control chip, which needs to work at about minus 270 degrees. Then there have to be detection equipment, quantum chips are made, there must be good equipment to detect whether it is working properly, just like we normally finish the chip, there must be process-oriented equipment to detect it, this Intel is also in cooperation with the cryogenic testing equipment factory, we have made the world's largest, can accommodate a whole 300mm wafer of the detection equipment, can simultaneously detect whether the quantum dots on it are in normal operation.
Observer: Is this one of the reasons why Intel sticks to the IDM (vertically integrated manufacturing) model?
Song Jiqiang: That's for sure, manufacturing silicon quantum chips requires too much accumulation in silicon process technology to do it.
Intel quantum chip Tunnel Falls
Observer Network: We know that many foreign companies set up departments in China are more inclined to sales functions, and Intel has set up Intel China Research Institute in China for a long time, I want to ask what role the research institute plays in the entire Intel system, and what role has it played in Intel's development in the Chinese market?
Song Jiqiang: Intel China Research Institute is part of Intel Research, and Intel Research is the engine of cutting-edge technological innovation in the entire Intel R&D system.
First of all, in the entire Intel R&D system, there are departments specializing in product research and development, they are based on relatively mature, can do industrialization, commercialization of technology, in one to two years into products to the market, or directly to support customer product and solution needs, this will account for a large part of the entire R & D system, such as the Intel Asia-Pacific R & D center in Shanghai, and the R & D team in Beijing, most of them are doing this work.
Intel Research is not directly to customers, but to the various product departments in Intel or the product development departments just mentioned, providing them with some technical input. At the same time, there is also a Technology Development department in the production system in the Intel R&D system, and the underlying semiconductor technology related to the process and materials is the main responsibility of the TD department. This department and Intel Research together form a relatively large research team. These two departments will issue papers and apply for patents, and will also do a lot of early experiments, TD departments will do research biased towards manufacturing, such as circuits, 3D packaging and other early experiments, Intel Research will also cooperate with TD departments.
However, Intel Research will be more inclined to how to use these technologies, such as hardware architecture such as processors, some power devices, power management devices, wireless transmission and RF devices, and some are in the software system, specializing in the underlying software optimization for hardware, there are also some virtualization, resource scheduling software for the cloud, and then there are artificial intelligence algorithms, artificial intelligence composition of special systems and other research.
In general, in the Intel Research system, it will be more related to some future different application fields, as well as some things to be done at each layer in the entire IT system stack and layer, because the technology development department itself is only doing semiconductor level research, but many layers above are done in the Intel Research Institute, including neuromorphic computing, which is a new computing architecture, which is all Intel Research is doing.
So, Intel Research is in the early stages of product development. Our input actually comes from many relatively mature technologies, such as universities and research institutions. We will establish some cooperative relations with universities and research institutions in various places to understand what kind of scientific research cooperation they are doing, and then import relatively mature and Intel-related technologies into Intel's system. This is also the primary responsibility of Intel Research, working with universities and research institutes. At the same time, we will also look at some government-level projects, because many times government-level project orientation is more important, just like the Chinese government once set the orientation, it will promote a wave of technological innovation.
In addition, there are relatively large research organizations of corporate partners, which is also a point for us to get some future directions. After having so many inputs, we go through the "three-stage" process of "Seek", "Solve" and "Scale" in the institute, to gradually lead it to the product department, and finally form a technology transfer. Therefore, Intel Research can be understood as an organization that screens some technologies from a relatively large pool that can be used for productization, and then through making some prototypes, and finally introduces it into the product development process, so we call ourselves "Industrial Research Institute", not pure papers, to measure the final technology transformation effect.
Observer Network: For example, the research and development of chips such as Intel 3 is mainly responsible for TD departments?
Song Jiqiang: That's right.
Observer Network: Neuromorphic computing and quantum computing, which are more cutting-edge technologies, are led by Intel Research?
Song Jiqiang: I will further clarify this problem, for example, neuromorphic computing, which is an innovation mainly at the chip architecture level, plus the construction of some software development kits above, so this part is basically all done by research institutes, but it will use Intel's latest process that can test chips. For example, its first generation is using a 14nm process, and the second generation is using an Intel 4 process, so it will continue to iterate with the new process.
But quantum computing is different, it is actually from scratch, and the institute and TD are working closely together. Here, for example, TD is mainly responsible for studying how to form qubits, including how to improve the yield of the future production process of qubits, how to correct errors, is they are mainly responsible, the research institute is responsible for how to control these qubits through radio frequency at low temperatures, how to construct some software stacks above, to test it, these two pieces are very tightly coupled together.
(The Observer Network has successively published two versions of exclusive interviews with Song Jiqiang: 1. Interview with Intel executives: China will definitely develop very well in RISC-V in the future; 2. Interview: In addition to using the most powerful lithography machine, what are Intel's tricks to surpass TSMC? ）