The Basis For Dialogue Between GIGALIGHT And Network Architecture
Currently, the networks we are communicating and constructing on do not possess a low-latency foundation. The belief in networks based on a 100G PAM4 underlying structure has been ingrained for many years. Although history cannot be reversed, it is necessary for us to question the networks, forming a theoretical basis for cognitive retrospection and moving forward anew.
We believe that not all networks are based on common infrastructure purposes. The diversification of network purposes is the ecological foundation for network diversification. Precisely because of this diversification, we cannot blindly accept the theory that networks based on a 100G PAM4 underlying architecture are infallible. Further refinements to this theory are inevitable.
Generally speaking, for applications like high-frequency trading and edge networks that require prompt responses, a more stable network underlayer is necessary. This underlayer does not necessarily rely on 100G PAM4 technology, as the stability of such networks is best achieved on the signal foundation of NRZ and 50G PAM4. If our goal is to achieve higher signal sensitivity and eliminate noise interference, NRZ and 50G PAM4 serve as the optimal bases for network practices. NRZ has an advantage over PAM4 primarily in signal-to-noise ratio (SNR), which is the foundation for eliminating DSP. As we all know, DSP introduces power consumption and significant network latency. Although both 50G PAM4 and 100G PAM4 belong to the PAM4 technology paradigm, their impacts on networks are strikingly different, with vastly different system SNR and power consumption implications. Technologies similar to LPO, which are currently trusted, can indeed form the basis for the next-generation network architecture akin to 50G PAM4. Attempts at this network have actually given us glimpses of the facts.
For high-frequency trading and edge networks, GIGALIGHT offers two evolutionary blueprints for technological practices. One is a network built on 400G QSFP-DD PSM8 optical transceivers, and the other is a network built on 400G QSFP-DD 2×FR4. These two networks can fulfill the purposes of high-frequency trading centers and edge networks, and they possess long-term effectiveness. If people must evolve towards 800G, we can still form a network architecture based on 16×50G PAM4. However, we also believe that both 50G PAM4 and 100G PAM4-based linear LPO and CPO technologies are goals for this network. The main dividing point remains: you must create a sufficiently stable, low-latency, and reliable foundation for this network. There is no need to pursue industry hotspots at the expense of blurring the practical functions of the network.
It is now believed that NV provides the most advanced network hardware for data center computing power, such as pluggable networks based on 200G SERDES/1.6T. However, this network faces challenges such as thermal power consumption and signal integrity. The industry has various technical concepts for improving this network, which are currently being tested and deployed. However, technologies like 3nm DSP, LPO, LRO, or immersion liquid cooling all require a certain amount of time for trial and error. Therefore, generally speaking, CPO may be a better solution. However, the premature definition of CPO makes this vision difficult to achieve in the short term. It is evident that NPO technology is the best practical blueprint for CPO implementation. Personally, I believe that data centers for computing power can continue to innovate along the path of 100G PAM4 without getting caught up in extending or evolving single-channel baud rates to 200G or 400G. Increasing baud rates reduces the system’s SNR and thermoelectric stability. I believe there are limits to deploying higher baud rate networks, and this model can be scientifically refuted, requiring reconsideration. Mainly, people must dispel the misconception that higher baud rates equate to technological advancement.
GIGALIGHT has a rich product line for AI&DC 800G computing power networks. We are currently establishing an innovative and differentiated product line for computing power networks starting from 800G and 1.6T. Products like GIGALIGHT’s 800G CWDM8/LR8 will serve as a supplementary architecture for 800G networks. For 1.6T, GIGALIGHT will begin a new product line layout and will release it by Q4 2025 at the latest.
The stagnation in 100G PAM4 technology research and development has led to difficulties in the innovation and evolution of metropolitan telecommunication networks. It is hard to believe that vast telecommunication networks in different locations can be built on the current underlying foundation of 100G PAM4 Ethernet. Future telecommunication networks must be based on three assumptions: first, building an infrastructure based on space wavelength division optics and a 50G network underlayer; second, developing a new technological platform for the next generation of 100G PAM4 DSP that is suitable for telecommunication networks; and third, establishing the next generation of high-speed networks entirely on the descent of various wavelength coherent technologies. These network architectures are not mutually exclusive, and each network can accommodate these three technologies. The main argument is that at different network nodes, corresponding technologies based on correct understanding are needed. However, this “correct understanding” must be based on the talent, preferences, and understanding of network rights of the network architects.
Based on the above understanding, GIGALIGHT will build the next generation of telecommunication network products within its own technological cognition. Indeed, this is a historical period dominated by computing power networks. However, people should not forget that it is the innovation in telecommunication networks that brings about the foundation for economic and industrial inclusiveness.
As for the most popular data center networks currently, since their concept is entirely focused on the general purpose of “data storage and retrieval,” this purpose is primarily based on simplicity and low cost. Therefore, their evolution is mainly driven by the interest in technological discoveries, the advanced deployment goals of the company’s future business, and assessments of its own economic strength and capabilities. Since the increase in general network baud rates cannot bring any additional benefits, we previously believed that 200G networks represented the optimal balance between cost and technology, and we still believe this today. Investments in 400G, 800G, and 1.6T general networks are mainly aimed at establishing a hierarchical structure within the network. If not driven by the demand for a hierarchical structure, the deployment of higher baud rate networks is solely motivated by the advancement of silicon photonics technology.
In the field of general data center architectures, GIGALIGHT has a comprehensive product line of silicon photonics, as well as a rich product line based on copper cable interconnection, VCSEL-based AOC, and EML technology. Due to the solidification of architectures, these products mainly compete on cost in the industry.
0 1
