Lightelligence has released its latest high-performance photonic computing processor PACE (PhotonicArithmeticComputine, photonic computing engine). However, the continuously growing model is clearly limited by the underlying computing power, limiting the further development of artificial intelligence.

Moore's Law has been in decline for nearly 50 years as nanotechnology continues to advance. The vigorous development of emerging fields such as artificial intelligence, 5G, and the Internet of Things has driven the explosive growth of global data. The demand for computing power is far higher than the growth in computing power supply predicted by Moore's Law. Traditional electronic chips can only be increased by increasing area and power consumption. To complete more calculations, it is gradually unable to meet the increasing data processing and energy saving requirements.

In the quest to go beyond Moore's Law, advanced craftsmanship has come into play. At the same time, due to its characteristics of high throughput, low latency, and low power consumption, using light instead of electricity to solve part of the calculation is also one of the ways to break through the existing bottleneck. Photonic chips that previously existed only in labs have recently made new progress.

Recently, Lightelligence released its latest high-performance photonic computing processor PACE (PhotonicArithmeticComputine, photonic computing engine).

Dr. Shen Yichen, founder and CEO of Xizhi Technology, said: The release of PACE is a milestone: it successfully verifies the advantages of photonic computing and provides a new development path for the integrated circuit industry.

Xizhi Technology was established in 2017 . Since its establishment four years ago, the company's total financing has exceeded 1 billion yuan. It has offices and laboratories in Boston, Shanghai, Hangzhou, Nanjing and other places. There are nearly 200 employees worldwide, and more than 100 employees in China. The core R&D team is from MIT, and 70% of the chip designers have more than 10 years of semiconductor experience.

In 2017, Dr. Shen Yichen published a cover paper in the journal Nature Photon as the first author and corresponding author, pioneering a new approach to photonic artificial intelligence computing. It is precisely because of this paper that more than a dozen startups have been established.

In April 2019, Xizhi Technology launched the world's first AD22286 photonic chip prototype board, successfully integrating the entire photonic computing system, which occupied half of the laboratory at that time, into a conventional-sized board, verifying that photons were used instead. Pioneering idea for high performance computing with electrons. At that time, 100 photonic devices were integrated on the prototype board, and the operating system clock was only 100kHz.

Two years later, the PACE released this time integrates 10,000 photonic devices, and the operating system clock reaches 1GHz. How do they span orders of magnitude performance gains?

Three Bottlenecks of Electronic Chips

Since 2012, the size of neural networks and computational models has exploded, with computational models doubling in size every 3 to 4 months on average. However, the continuously growing model is clearly limited by the underlying computing power, limiting the further development of artificial intelligence.

Dr. Shen Yichen believes that the development of electronic chips has encountered three main bottlenecks: computing power, data transmission and storage. Among them, the computing power bottleneck mainly comes from Moore's Law failure, as well as power consumption and heating problems caused by the process approaching the physical limit.

As transistors get smaller, electron tunneling across transistors becomes more severe, so even with smaller transistors, the power consumption of a single transistor cannot be reduced further. Under this premise, the industry has two solutions, single-chip area increase or multi-chip interconnection.

However, as the area increases, longer copper wires are required for data transmission, and the heat loss of the copper wire is proportional to the length, that is, the larger the chip area, the greater the heating and the higher the power consumption.

Likewise, there are some issues with multi-chip interconnects. First, the interconnection bandwidth between chips is limited, that is, the interconnection efficiency is low. Second, copper wires still add power to the system. For example, after 100 chips or boards are electrically connected, the computing power may only be increased by about 10 times that of a single board.

Therefore, Dr. Shen Yichen believes that light is the most suitable underlying technical method to solve these difficulties. First of all, in terms of data processing, light has fully proved its leading position and advantages in the field of optical communication. Currently, all long-distance communication, including data centers, data between servers and servers is carried out with fiber optics instead of copper wires. We also believe that light entering the chip to help the operation is an inevitable direction.

Three technologies of Xizhi photonic computing

As mentioned earlier, light is the underlying technical method to solve the three bottlenecks of the current electronic chip computing power, data transmission and storage. From the perspective of applications such as big data and artificial intelligence, more and more computing power requirements come from linear computing, and the efficient linear computing method invented by Xizhi is one of the important advantages of optical chips.

Xizhi divides its technology into three parts: OMAC (Matrix Multiply Accumulation Computing), ONOC (Optical Network) and Optical Network. According to Dr. Shen Yichen, OMAC is a kind of analog computing, which uses optical analog signals to replace traditional electronics for data processing. Electronics perform data processing. Data can be loaded on the intensity or phase of light and interfere with each other through propagation in the waveguide. The main implementation method is to use a silicon photonics process platform compatible with the current electrical chip fabrication process CMOS, and use optoelectronic collaborative design to perform optical matrix multiplication.

The advantage here is that, firstly, the optical matrix multiplication is more parallel and can be computed at higher throughput. At the same time, its energy efficiency is comparable to or even better than current electronic chips, because light does not heat up as it travels. In addition, less time is required to complete the matrix operation, i.e. the delay is much lower than that of electronic chips. Finally, the silicon photonics process has fairly low requirements on the process technology. For example, CMOS process lines of 65 or 45 nanometers can meet all the requirements of current optical chips and optical computing. The future technology iteration of silicon photonics does not require special requirements for the process, but iterates technology from other aspects such as main frequency and wavelength.

ONOC, or Film Lighting Network, mainly transmits data on films by replacing copper wires with waveguides, including realizing optical communication between films. There is also the communication of the large-chip lighting bus, a fixed communication network topology is constructed on the optical chip, and the data interaction based on the movie lighting network is realized through the optical connection. Finally, using some wave means to propagate data, the advantages are greater bandwidth, lower energy consumption, latency will be much better than copper wire, and it is not sensitive to distance.

The last inter-chip optical network will be further extended to multiple boards and more servers. Chips and chips are directly connected by optical fibers, and data between chips is transmitted by light.

The world's only demonstration of photonic advantages, PACE exceeds 3080 -ic/" title="3080 model">3080-ic/" title="3080 model">3080-ic/" title="3080 model">3080-ic/" title="Model 3080">3080-ic/" title="Model 3080">3080-ic/" title="Model 3080">3080-ic/" title="Model 3080">3080-ic/" title= "Model 3080">3080-ic/" title="Model 3080">3080-ic/" title="Model 3080">3080-ic/" title="Model 3080">3080-ic/" title="3080 Model">3080-ic/" title="Model 3080">3080-ic/" title="Model 3080">3080-ic/" title="Model 3080">3080-ic/" title="Model 3080" >3080-ic/" title="Model 3080">3080-ic/" title="Model 3080">3080-ic/" title="Model 3080">3080 Hundred Times!

Xi Zhi believes that the most important technological evolution point of optoelectronic hybrid computing is to continuously increase the integration of devices on a single optical chip. In fact, from the earliest 4x4 multipliers to 64x64 multipliers, to the current 2.5D optoelectronic hybrid packaging, Xizhi realized the integration of 10,000 optical devices on one chip within four years.

Due to the breakthrough in integration, PACE is the latest operational computing processor that West Intelligence can currently demonstrate. It is the most integrated photonic chip known in the world and the first computing system in the world to demonstrate the advantages of photonics. It could be orders of magnitude better than current electronic chips in some promising algorithms.

So what areas can the advantages of photonic computing be reflected in? NP-CompleteProblem (Multiple Complexity Uncertainty Problem, NPC) can be said to be the most difficult mathematical problem to solve in the world, such as the prediction of protein structure in biological information, logistics and transportation scheduling, chip design, material research and development, etc. However, at present, NPC does not have multiple algorithms, and can only use the exhaustive method to test them one by one, and finally get the answer. But if it can effectively solve one of the problems, it can also be mapped to the other problems.

Due to the characteristics of photonic chips, PACE can achieve low latency through repeated matrix multiplications and clever use of tight loops composed of controlled noise, so PACE can compute NPC problems hundreds of times faster than GPUs. Therefore, PCAE has greater commercial application prospects in solving NPC problems.

It is understood that compared with NVIDIA RTX3080GPU, PACE takes less than 1% of the time required to run the same recurrent neural network algorithm simultaneously.

Photoelectric hybrid structure based on existing ecology

In fact, the structure of PACE consists of an optical chip and an electrical chip. Electrical chips are mainly used for data storage and digital-analog hybrid scheduling, and optical chips are mainly used for data computing. It can be understood here that the optical chip is only the underlying hardware support, and the electrical chip related to information conversion and software is used for digital processing. All instructions, compilation and software will be on digital electronic chips. Therefore, like the existing digital chip ecosystem, optical chips are only replaced at the underlying computing side.

As the design of optoelectronic hybrid chips, some people may worry that the process is difficult to mass-produce. In fact, Shen Yichen said that silicon photonics chips use CMOS technology, which can solve 90% of the core problems. Since the silicon-based CMOS process is basically used, mature software such as electrical and thermal including simulation can be used directly.

At the packaging level, PACE uses chip stacking, a 2.5D and 3D packaging scheme similar to HBM. The only difference at present is that an interface needs to be added to the packaging scheme to import the light source into the optical chip.

How far is the commercialization of optical chips?

When talking about the commercial prospects of this technology, Dr. Shen Yichen emphasized to reporters that optical computing is not just about optical chips. In the foreseeable future, it will be optoelectronic hybrid computing deeply integrated with electronic chips. Compared with electrical chips, optical chips are more processors that undertake the main tasks, mainly linear computing and data networking. However, one benefit of electronic chip instructions is that it is compatible with the existing market environment and software environment.

In addition, it should be noted that Xi's smart optoelectronic hybrid chips cannot be used in fields familiar to consumers, such as PCs, mobile phones, codec chips, etc. At the same time, this is not the scope of Xi's intelligent technology. When selecting application scenarios, Xi intelligent technology will first cut into big data, including cloud computing, intelligent driving, financial quantitative transactions, biological drug research and development and other scenarios.

Shen Yichen said that as a disruptive technology, it must go through a long commercialization process. He revealed that in the first stage, that is, within one to three years in early 2022, application scenarios with particularly strong computing power and delays will be implemented, including finance, large-scale cloud services, optimization of non-artificial intelligence, and high-performance computing.

In the second stage, with the implementation of the product, after reflecting the advantages of optical computing in different application scenarios, it will invest in a larger-scale team for the artificial intelligence training market.

Xizhi will extend to GPUs in the third phase, including markets such as in-vehicle chips.

These are what we think is a huge demand for computing power, but we need more mature hardware, software systems and further markets. Therefore, Shen Yichen believes that the commercialization of technology will be a long process, which requires continuous changes and different application scenarios and industries.