GPT-4o Drives Computing Power Trends: Optical Communication Technology Heats Up
On May 13, 2024, OpenAI launched its new flagship model, GPT-4o, capable of real-time reasoning across multiple formats. This advancement represents a significant leap forward in natural human-computer interaction, allowing the processing of any combination of text, audio, and images, and generating any combination of text, audio, and images as output. The “o” in GPT-4o stands for “omni,” highlighting its multi-modal capabilities. The audio input response time of GPT-4o is just 232 milliseconds, with an average of 320 milliseconds, which is similar to the response time of human conversation. Its text and coding capabilities are on par with GPT-4 Turbo, while non-English text functionality has seen significant improvement. The application programming interface (API) is faster and 50% more cost-efficient compared to its predecessors. GPT-4o stands out, especially in visual and audio comprehension, surpassing previous models.
Previously, voice mode in ChatGPT models had an average delay of 2.8 seconds (GPT-3.5) and 5.4 seconds (GPT-4). These voice modes consisted of a pipeline mechanism with three independent models: one simple model converting audio to text, then GPT-3.5 or GPT-4 processing the text, followed by another simple model converting the text back to audio. This mechanism couldn’t capture tone, multiple speakers, or background noise, leading to the loss of significant information such as laughter, singing, or emotional expressions.
GPT-4o represents a new, fully end-to-end AI model trained in text, visual, and audio formats. It integrates all existing modes, with all inputs and outputs being processed by the same neural network, though its capabilities and limitations are still being explored.
Lifecycle of AI Models like GPT-4o
AI models, particularly large models like GPT-4o, undergo several stages throughout their lifecycle:
- Development Phase: This phase involves extensive experimentation to determine model features (parameters) and effective algorithms.
- Training Phase: A large volume of data is input into the model, allowing it to learn and create intrinsic data structures through machine learning.
- Fine-Tuning Phase: After the initial training, the results are refined and adjusted to improve the model’s performance.
- Deployment Phase: Once deployed, the model handles real-time user demands, processing vast amounts of data.
AI models belong to the category of High-Performance Computing (HPC), which requires powerful, reliable, and controllable computing capabilities across all stages to ensure low-latency and high-speed performance. Given the large data volumes required for video training, the growing development and usage of AI models will drive increased demands for upgraded computational power.
Meeting AI’s Training Needs with Enhanced Computing Infrastructure
To meet the massive computational needs of AI models, especially for industries like transportation, healthcare, education, energy, and finance, the 2024 China Mobile Computing Power Network Conference showcased the new integrated computing power network infrastructure. The conference introduced three independently controlled intelligent computing centers (Harbin, Hohhot, Guiyang), with a total scale of nearly 60,000 GPUs. Additionally, nine other intelligent computing centers were launched, offering a combined total computing power of 11 ExaFLOPS. The Hohhot Intelligent Computing Center is currently the largest single liquid-cooled data center operated by global telecom operators, which has garnered widespread attention.
At the core of these computing centers is the OTN-based (Optical Transport Network) computing optical network, which supports these high-performance computing networks. The OTN network is built on optical fiber cables for high-reliability, high-quality connections, and uses optical signals for ultra-high bandwidth transmission and widespread connectivity. It also features a highly coordinated, intelligent management system for security and control.
Advances in Optical Communication Technology
The rapid growth of intelligent computing centers/data centers is pushing for increased OTN system transmission capabilities, which are driven by the performance of coherent optical modules. This requires upgrades across various levels of the network:
- Top-of-Rack (TOR) to Server: The speed has increased from 100G (AOC/DAC) to 200G (AOC).
- Leaf (including leaf switches) to TOR: The speed upgrade is from 400G Q-DD (SR8/DR4) to 800G (PSM8/4).
- Spine (including routing and forwarding) to Leaf: The speed has doubled from 400G Q-DD (DR4/FR4) to 800G (PSM4/FR4).
- Data Center Interconnection (DCI): The speed has moved from 400G Q-DD (ZR) to 800G (ZR).
The increasing business traffic has driven the need for continuous upgrades to the OTN backbone network’s port speeds. Starting from the commercial use of 100G in 2012, then 200G by 2016, and 400G wavelength division multiplexing appearing by 2020, the market is now testing 800G wavelength division multiplexing. By 2025, the growth rate for 400G will slow down, while 800G+ port speeds are expected to surge. By 2027, 800G+ ports will surpass 200G, becoming the mainstream system rate.
Applications and Future Trends in 800G Optical Modules
The 800G optical modules, based on single-channel rates, are typically classified into two types:
- 100G × 8 single-channel modules: Simpler in design.
- 200G × 4 single-channel modules: Require higher performance for optical components and need variable-speed conversion.
These 800G optical modules are expanding in applications such as:
- Data Center Interconnect (DCI): Ensuring seamless connections between data centers.
- High-Performance Computing (HPC): Facilitating fast data transmission, reducing latency, and optimizing overall system performance.
- 5G/Telecom Networks: Supporting next-generation communication architectures to meet the demands of next-gen networks.
The widespread adoption of 800G modules and infrastructure will play a crucial role in the expansion of AI models and the increasingly complex demands of industries worldwide.
Conclusion: The Future of AI and Optical Communication
The development of models like GPT-4o is leading the charge in creating more intelligent, multi-modal AI systems. As the need for large-scale training and real-time processing increases, so does the demand for more advanced computing power and infrastructure. Optical communication technology is at the heart of this transformation, providing the necessary speed, reliability, and bandwidth for handling the vast data streams that these AI systems generate.
The future of AI-powered industries will rely heavily on advancements in both computational power and the underlying optical network technology that supports them.
