- All
- Product Management
- News
- Introduction
- Enterprise outlets
- FAQ
- Enterprise Video
- Enterprise Atlas
400G optical module, is it ChatGPT enough?
Release time:
2023-10-18
At the beginning of this year, the AIGC application represented by ChatGPT was born, and the degree of popularity swept the world. Domestic Internet companies are also competing for large-scale large-scale model investment. One of the problems caused by this is the interconnection interface and bandwidth of GPU. Most of the common Ethernet architectures previously adopted in China are based on 400G system. The difference between the third generation and the second generation 400G is that the host board side of the former is based on the next generation 112G Serdes/PHY, which can be upgraded to 800G system very flexibly.
Therefore, at present, large domestic Internet companies mostly use 800G interface to reduce the speed to 400G in the interconnection of AI large models, but in the subsequent interface applications, they are actually based on 112G technology.
At present, 400G can still meet the needs of AIGC, but with the increasing demand of GPT-4 and GPT-5 for computing power, it is estimated that the next generation of GPU interconnection will need more than 800G bandwidth to meet the demand, which will also push the whole industry chain to rapidly shift to the next generation of 112G-based new architecture and multi-communication model.
Today, not only the internal communication of the board, but also the connection between the boards and the switch, have adopted this typical structure. In the case of a sharp increase in connection demand, how to ensure that the multi-channel 112G can operate stably under different circumstances is a problem that the entire industry chain is trying to solve.
From 400G to 800G, what should I pay attention?
Judging from the switching process, what are the points that need special attention from 400G to 800G?
The AI large-scale model is a whole system, covering chips, boards, servers, and switches, and ultimately constitutes a back-end management and control system that requires corresponding resource allocation. In terms of optical communication, the primary consideration is optical connection technologies such as optical modules and IOC cables. The distance involved in these optical connections may range from 3 meters, 10 meters or even 100 meters, and the typical interconnection methods are different.
Therefore, when switching from 400G to 800G systems, there are several core points worth noting:
First of all, we should pay attention to the packaging of optical modules. Silicon light is a very important technology for the next generation of AI applications, so the current optical module design is more performance-oriented. At present, silicon light is a single-mode technology, but with the development of technology, it may eventually be the same as the traditional multi-mode light technology in terms of cost, and the resulting integration will be higher.
Second, digital signal processing (DSP). Although there are some linear drive pluggable optical module (LPO) technology, the mainstream application is still based on DSP technology, and there are still different views in academia and industry on the performance balance between the two. Some voices think that DSP and LPO have reached a good performance balance and can be applied on many switches to achieve better performance. Another part thinks that LPO technology performs very well in the first 10 minutes, but as time goes by, Margin will get lower and lower, directly leading to thermal noise. Without DSP, the effect of reflection or temperature will increase the pressure on the back-end main chip processing.
If the balance between DSP and LPO can be achieved, it can ensure the long-term stable operation of the switch under different environmental conditions.
Finally, 112G technology maturity. This involves the entire industry chain from plug-in, PCV, advanced packaging and switch performance, system integration and processing, and now every step seems to be facing huge challenges.
We are discussing 800G today, and we may discuss 1.6T next year. This is the acceleration brought by the changes in data traffic caused by the advanced deployment or large-scale deployment of AI large models. Optical module manufacturers must lay out in advance.
Next
Next
Latest information
Optical Modules Lead the Future of Data Centers
The wind of the big model has been blowing for more than half a year, and the arms race triggered by the GPT-4 has not stalled, but has intensified. Many people pay attention to the emergence of intelligence, and some people pay attention to computing power. In fact, the optical communication industry is also driven by AI.
2023-10-18
All rights reserved©: Shenzhen Opticking Technology Co., Ltd.
Headquarters and Marketing Center:1701-1714, Building C, Jinxiu Phase III, No. 85, Hudi Pai, Songxuan Community, Guanhu Street, Longhua District, Shenzhen China
Telephone:0755-83425935 Fax: 0755-21002709