New paradigm of optoelectronic hybrid computing = optical computing + optical interconnection

Integrating hundreds of millions of transistors into a chip the size of a fingernail and increasing its integration density has been the main method for increasing the computing power of chips in the past few decades, and it is also the core of Moore’s Law, which has led the industry for more than half a century. content. However, due to the rapid development of artificial intelligence, big data, Internet of Things and other industries, the wave of digital economy is sweeping, and the demand for computing power as a core productivity has surged, which gradually conflicts with the physical limit of the chip itself. Moore, who was once regarded as the “golden rule” The law is facing the dilemma of failure.

Light or will become a breakthrough to solve this problem? Photons have the advantages of high throughput, low latency, low energy consumption, and are less susceptible to changes in temperature, electromagnetic fields, and noise. Previously, photonic technology was often used in the field of long-distance communication transmission, and optical fiber communication has become the main transmission method of various communication networks. However, the role that light can play in the progress of human society may be far greater and more important than we imagine.

Xizhi Technology took the lead in turning its attention to the field of light, and proposed the pioneering concept of “New Optoelectronic Hybrid Computing Paradigm”, trying to provide the integrated circuit industry with a new “problem-solving idea” that is different from the current traditional chip computing paradigm. Recently, Dr. Meng Huaiyu, CTO of Xizhi Technology, gave a detailed introduction to this concept at the “DeepTech Technology Innovation Week Advanced Computing Forum” with the theme of “New Paradigm of Optoelectronic Hybrid Computing under the Explosion of Computing Demand”.

A new revolution starting with optical computing

The so-called light calculation refers to the use of the physical properties of light to complete linear calculations. Dr. Huaiyu Meng pointed out the three advantages of optical computing by taking optical computing, which is common in life, as an example.

The first is low latency, the time required for the observer behind the glasses to perceive the image change in front of the glasses is equal to the time it takes to travel this distance at the speed of light—almost insignificant; and the second is low power consumption, where the glasses themselves do not consume energy , all the energy is consumed in the generation and absorption of the optical signal; the last is high flux, when the optical signal changes at a high speed, the information received behind the glasses also changes at a high speed, that is, the two-dimensional Fourier transform of the glasses is undergoing high-speed high-throughput computing.

Although the glasses are not programmable, their principles have inspired the realization of optical computing. In order to use light to realize a programmable and truly useful computing system, Dr. Shen Yichen, founder and CEO of Xizhi Technology, pioneered a new computing architecture that uses integrated photonic technology to realize deep learning, and founded Xizhi Technology in 2017. In 2019, Xizhi Technology released the world’s first photonic chip prototype board, which successfully verified the pioneering idea of ​​replacing electrons with photons for high-performance computing. In 2021, on this basis, the Xizhi Technology team released the high-performance photonic computing processor PACE (Photonic Arithmetic Computing Engine, photonic computing engine), through repeated matrix multiplication and clever use of tight loops composed of controlled noise to achieve low delay, resulting in a high-quality solution to the Ising problem.

Dr. Meng Huaiyu said: “PACE mainly takes advantage of the low latency of optical computing. It can complete a single iteration of the Ising problem within 3 nanoseconds, and the speed is more than 800 times that of current high-end GPUs.”

Comparison of PACE and current high-end GPU performance

Optical interconnection, the other half of the new paradigm of optoelectronic hybrid computing

The other half of the “new paradigm of optoelectronic hybrid computing” is to solve the data interconnection problem, that is, the “memory wall” problem, which mainly includes two parts: capacity and bandwidth. Today, with the explosion of computing power, the growth rate of the corresponding hardware is unmatched. Taking Transformer, a typical AI model, as an example, the fact that the size of the algorithm has increased by 240 times in two years is the fact that the hardware storage capacity has only increased by 2 times. Hence the capacity bottleneck of the memory wall, which is how to accommodate larger applications.

Another big challenge is bandwidth bottlenecks. Dr. Meng Huaiyu explained that if the chip is imagined as a flat block, the computing power of the chip is proportional to the area of ​​the block, and the external bandwidth of the chip is proportional to the side length. Therefore, when the density of transistors on the chip is getting higher and higher, if the edge length density of the chip is increased by 2 times, the computing power density will be increased by 4 times. So, whether it’s Moore’s Law moving forward, or new computing paradigms that increase computing power per unit area, the more bandwidth required to “feed” computing power will become a problem. In the past 20 years, the computing power of hardware has increased by 90,000 times, but the DRAM bandwidth and network bandwidth have only increased by 30 times.

The solution given by Xizhi Technology is a new paradigm of data interconnection – optical interconnection. Compared with the performance of electrical interconnection, which gradually decreases with distance, optical interconnection is much less affected by distance. Dr. Meng Huaiyu said: “Ideally, for data transmission over 10mm, it is more advantageous to use optical interconnection, which can bring greater possibilities to solve bandwidth bottlenecks and capacity bottlenecks, which is also Xizhi Technology’s new paradigm of optical interconnection. underlying logic.”

Optical interconnection and electrical interconnection comparison

At present, the optical interconnection solution has been applied in the data center, but because the distance between the optical module and the digital chip using the optical module is often more than 1 meter, the optical interconnection is limited by the electrical interconnection, which leads to the application scope of the optical interconnection. It is limited between racks, within racks, and even within servers, using very little optical interconnection. In order to eliminate the bottleneck of electrical interconnection, the new paradigm of optical interconnection advocated by Xizhi Technology is to highly integrate photoelectric conversion and digital chips to form “chips that emit light”, thereby broadening the feasibility of many computing paradigms. Dr. Meng Huaiyu took the current trend of “resource pooling” in data centers as an example. He said: “I would interpret this large-scale resource pooling as a ‘sharing economy’ of computing resources. Currently, one server needs to access another server. It will be more difficult to use the resources of the Internet because of their poor interconnectivity. Optical interconnection can help achieve better interconnectivity and make it possible to share resources on a large scale. Finally, through resource pooling, we can make each computing chip All have access to larger memory and have larger bandwidth, thus solving the memory wall problem.”

Optical interconnection makes data center “resource pooling” possible

Ultra-large-scale optoelectronic hybrid integration is the underlying technology that enables all of the above. In this regard, Xizhi Technology has also completed relevant technical verification, successfully stacking an integrated silicon photonics chip and an electronic chip vertically in a 3D package, so as to minimize the distance between the two chips. Transceiver has more than 1000 times higher integration density.

Finally, Dr. Meng Huaiyu also introduced Xizhi Technology’s “optical hybrid wafer-level computing platform” solution. Today, many companies in the industry have introduced the concept of “wafer-level computing platforms”, that is, to achieve higher performance through larger chip areas, such as Cerebras’ WSE chips. But their limitations are also obvious: first, because electricity is not suitable for long-distance communication, only nearest neighbor data transmission can be performed; second, they will face a more serious “memory wall” problem. In this regard, the “Optoelectronic Hybrid Wafer-Level Computing Platform” solution achieves any interconnection topology, low latency and low energy consumption through wafer-level on-chip optical interconnection. At the same time, in order to break the “memory wall”, a remote resource pool can be set up, and through high-efficiency optical interconnection, it can directly access the optical network inside the wafer-level computing platform, and finally realize the optimal configuration of all computing resources.

Xizhi Technology “Optoelectronic Hybrid Wafer-Level Computing Platform”

Since its establishment in 2017, Xizhi Technology has been committed to providing customers with more creative and efficient computing power support through a new optoelectronic hybrid computing paradigm composed of optical computing and optical interconnection. As of March 2022, Xizhi Technology has raised a total of more than 1.4 billion yuan, leading the global photonic computing track. Its latest solution based on optical interconnection is currently under continuous research and development and will be officially released in the form of products in the near future.