Since chip manufacturers began to take the chip microfabrication process as the focus of development, the discussion in the industry about whether Moore's Law is coming to an end has not stopped.
Xu Juyan, an academician of the Chinese Academy of Engineering, once listed four types of technology directions for CMOS and new devices, von Neumann architecture and emerging architecture on the prospective topic of the next wave of chip technology:
One is the "silicon von" paradigm that combines silicon CMOS technology with von Neumann; one is the "silicon-like" paradigm that can enter the same binary switch as CMOS, and the combination of new devices and von Neumann architecture; the other is the "brain-like" paradigm that still uses existing silicon CMOS device technology, but does not pass through the von Neumann architecture but through synaptic transmission and storage integration; and the last type is the "emerging" paradigm that is done with emerging architectures and new devices.
Light, which can be seen as one of the last emerging paradigms, has lagged behind in competition with electronic chips. But now there is a new breakthrough, with the potential to become a strong member of the post-Moore era.
This month, photonic computing chip company Xizhi Technology released its latest high-performance photonic processor, integrating more than 10,000 photonic devices in a single photon, running a 1GHz system clock, and running a specific recurrent neural network at speeds up to hundreds of times that of current high-end GPUs, which fully verifies the superiority of photonic chips.
Photon calculations that fail on a universal computing track
Since the performance improvement of the chip given by photonic computing is much greater than that of the electronic chip, why has the light that has been discovered by humans since the Mozi period failed to outperform the electronics in the chip field?
Shen Yichen, founder and CEO of Xizhi Technology
Shen Yichen, founder and CEO of Xizhi Technology, told Leifeng Network that in fact, when semiconductors were just starting out, there were two technical paradigms of light and electricity, but after a period of development, the entire industry almost covered the computing paradigm based on digital electronics.
The digital electronic computing paradigm has been able to win for two reasons.
On the one hand, the rise of Turing computing, including von Neumann's digital chip architecture, can allow digital chips to achieve almost all general-purpose computing through logic gates, and a wide range of applications.
On the other hand, in the 1980s, when the logic gate-based photon numerical computing and electronic computing competed on the general computing track, they were based on transistors to do calculations, but when the electrotransistors continued to shrink as the process advanced, the size of the optical transistors has not been able to break the wavelength limit, can not be smaller than 100nm, and has since lost the electronic chip.
However, in recent years, the situation has changed, especially in the development of electronic chips so far, in terms of computing power, data transmission and storage have encountered bottlenecks, continue to seek a breakthrough in the paradigm of electronic computing technology is difficult.
Computing power bottleneck is the most commonly talked about problem under the third wave of artificial intelligence, the electronic tunneling phenomenon brought about by transistor shrinkage, resulting in the transistor power consumption under the advanced process can not be further reduced, so some companies hope to increase the computing power by expanding the chip area, but in the end it is found that the larger area requires longer copper wires, longer copper wires generate more heat, so the chip energy efficiency ratio has not been greatly improved with the expansion of the area.
NVIDIA increases the computing power by means of electrical interconnection chips, but it is limited by the interconnection bandwidth, interconnecting 100 chips or boards can only reach 10 times the computing power of a single plate, and the interconnection efficiency is not high.
Now, the architecture of optical computing is changing. The popularity of AI computing has also brought broader prospects for optical computing. "We believe that light is the best underlying technical approach to solving these dilemmas." Shen Yichen gave his own answer.
Optical computing is completely different from electronic computing, it takes the information processing of photons as the carrier, relies on optical hardware instead of electronic hardware, replaces electrical operations with optical operations, and is good at rapidly parallel processing of highly complex computing tasks.
He believes that on the one hand, light in the field of communication has fully proved its superiority, all current long-distance communication, including communication between servers in the data center, are based on optical fiber transmission, on the other hand, more and more artificial intelligence to do linear operations, and the interference of light itself is linear engineering, compared with electricity has a natural advantage.
This answer has a strong argument, and in 2016, the research team of Dr. Shen Yichen of MIT created the first optical system, and in 2017, the results were published on the cover of the top journal Nature Photonics.
Its innovation has two, the hardware use of optical interferometer as a basic matrix operation unit to effectively replace the traditional electronic transistor, algorithm developed a series of algorithms that effectively reduce the amount of deep learning computation without sacrificing performance conditions and adapt to photonic chips.
The results of photon computation have begun to appear
The release of this experimental result inspired a large number of people around the world to start paying attention to photonic AI chips, and Shen Yichen himself also went from the MIT team to the industry and founded Lightelligence, a company that focuses on the research and development of photonic chip related technologies.
A year and a half after the company's establishment, Xizhi Technology successfully developed the world's first photonic chip prototype board, and successfully demonstrated the use of photonic chips to run Google Tensorflow's own convolutional neural network model to process MNIST data sets, that is, the use of a benchmark machine learning model for computer recognition of handwritten digits to process data sets. More than 95% of the operations of the entire model are processed in the photon chip, and the advent of this prototype board has proved to the world the feasibility of using photons instead of electrons for AI calculations.
The test results show that the accuracy of photonic chip processing has approached that of electronic chips (more than 97%).
Less than two years after the release of this prototype board, in December 2021, Xizhi Technology made new progress by releasing high-performance photonic computing processing - PACE (Photonic Arithmetic Computing Engine) - integrating more than 10,000 photonic devices in a single photonic chip, running a 1GHz system clock, running a specific recurrent neural network hundreds of times faster than current high-end GPUs.
"Compared with the prototype board launched in 2019, our PACE has increased the integration of photonic devices by about two orders of magnitude, from 100 photonic devices to 10,000 photon devices; the operating system clock has been increased by four orders of magnitude, basically reaching the clock of the current electronic chip." Shen Yichen said.
It is worth noting that PACE is not a pure photonic chip, PACE contains a 64x64 optical matrix, the core part of which is stacked upside down in a 3D package by an integrated silicon optical chip and a CMOS microelectronic chip. Among them, the electric chip is mainly used for data storage and digital-analog hybrid scheduling, and the optical chip is mainly used for data calculation.
"We think that electrical storage technology, especially in terms of high-speed storage reads, will be ahead of light for a long time. This is why we use optoelectronic synergy. Shen Yichen explained.
When PACE is running, each input vector value is first extracted from the on-chip memory, converted from a digital-to-analog converter to an analog value, and applied to the corresponding optical modulator through a micro-bump between the electronic chip and the photonic chip to form an input light vector.
Then, the input light vector is propagated through the optical matrix to generate the output light vector and reach a set of photodetector arrays, thereby converting the light intensity into a current signal.
Finally, the electrical signal returns to the electronic chip through microconvexes and the digital domain via transimpedance amplifiers and analog-to-digital converters.
Shen Yichen said that the optical modulator used by Xizhi Technology is a small-size high-speed adjustable optical modulator based on the Mahzende interferometer scheme to do the interference between light and light, and is optimized in coordination with the algorithm. Because the light does not exude heat when propagating, it takes less time to complete the matrix operation, the delay is lower than that of the electric chip, and the matrix multiplication parallelism ability is stronger.
PACE adopts iterative methods to solve mathematical problems that are difficult to solve efficiently in the world - polynomial complexity non-deterministic problems, involving protein structure prediction in bioinformatics, logistics and transportation scheduling, material research and development, etc., with broad commercial application prospects.
Xizhi Technology has also made its own product planning, planning to land in application scenarios with strong computing power and time delay pain points in 1 to 3 years from 2022, such as financial and cloud service providers, and then strengthen the layout of the training market, and finally extend to GPUs, in-vehicle chips and other markets.
"Breaking through" Moore's Law, completely replacing or developing in parallel?
In fact, in addition to Xizhi Technology, many large manufacturers have begun to invest in the research and development of photonic computing chips. It is understood that Huawei, Intel, and Nvidia currently have entering photonic computing.
Is the increasing number of players on the photonic computing track mean that in the future, photonic computing chips will have the ability to completely replace electronic chips and "break" Moore's Law?
Returning to the discussion of Moore's Law, although photonic computing is a very different technical route, silicon optical chips are still based on traditional CMOS processes, relying on existing ecology, firmware and software to meet basic design needs. In terms of process manufacturing, only minor modifications are required in the steps, such as the introduction of other new equipment in the manufacture of light detectors.
Previously, the light-based transistor was huge, and it was defeated in the competition with electronic computing, but now the photonic computing has changed the original path of competition between transistors and electronic computing, using the advantages in linear operation to do optical devices, the computing performance of an optical device is equivalent to thousands of electrotransistors, and the 65nm or 45nm CMOS process process can meet all the process requirements of existing optical chips, even in the future, silicon optical technology iterations will not be particularly stringent on the process requirements, more is the main frequency, Iterations in wavelength.
This means that even if Moore's Law is close to the physical limit, it will not have much impact on photon computing chips. However, this does not mean that photonic chips can completely replace electronic chips.
Shen Yichen said that in the foreseeable future range, it is a photoelectric hybrid operation that deeply combines photons and electronic chips, and the interaction between Xizhi Technology's optoelectronic hybrid chips and customers is done through electric chips, and all instruction compilers and SDKs are carried on the electric chips. Compared with the electric chip, the optical chip mainly carries the two major parts of linear computing and data network, and the advantage of the electric chip is that it is compatible with the existing market environment and software environment.
"The relationship between light and electricity is like that of new energy vehicles and fuel vehicles, but there are changes in the engine and battery, and devices such as tires are still the same, which will be effectively compatible with existing use scenarios." Light does not completely replace electricity. Shen Yichen said.