800G is not only a leap in speed, but also a comprehensive evolution of data center networks in the AI era. From PAM4 modulation to CPO packaging, from optical modules to liquid cooled heat dissipation, every step faces technological breakthroughs. This silent ‘speed and passion’ is reshaping the underlying rules of AI competition, and the layout of 800G has changed from optional to mandatory.
800G technology core: How to achieve transmission of 800 billion bits per second?
The popularity of high-order modulation technology PAM4: To achieve rate doubling with the same number of channels (such as 8 channels), each symbol must carry more information. NRZ (Non Return to Zero) encoding only carries 1 bit of information per symbol, while PAM4 can carry 2 bits. This means that 800G optical modules require more complex digital signal processing (DSP) chips to cope with higher signal-to-noise ratios and signal attenuation.
Competition in various packaging forms:
Pluggable optical modules (such as 800G QSFP-DD/OSFP): are currently the mainstream in the market, inheriting the maintainability and flexibility of previous generations of technology, but facing limits in power consumption and density.
Co encapsulated optics (CPO): Encapsulating the optical engine and switch chip together in the same slot greatly shortens the electrical channel distance, significantly reduces power consumption and latency, and is the ultimate solution for future ultra high speed networks.
Linear Drive Pluggable Optics (LPO): A compromise solution that eliminates DSP chips and relies on the linear analog characteristics of switch ASICs and optical modules to drive, aiming to achieve plug-in convenience and low power consumption close to CPO.
Deploying 800G: Full stack considerations beyond switches
Deploying an end-to-end 800G network is a systematic project:
Optical modules and fibers: 800G requires higher performance single-mode fiber (SMF) to support longer distance transmission, or multi-mode fiber (MMF) for short-range interconnection. SR8 (short distance), DR8 (500m), FR4 (2km) and other optical modules of different specifications need to be selected according to different scenarios such as inside the cabinet, across cabinets, and across data centers.
Switches and routers: Spine switches in core switches and spine topologies require 800G line cards with high port density, and their switching capacity must be at the petabits level to meet the requirements of non blocking switching.
Network card and server: GPU servers (such as NVIDIA’s Grace Hopper supercomputer) and DPUs (data processing units) are natively supporting 800G network interfaces, ensuring smooth paths for data from GPU memory to network cables.
Power supply and heat dissipation: The power consumption of an 800G pluggable optical module may be as high as 20-30 watts, nearly twice that of a 400G module. A fully loaded 800G switch rack will generate an astonishing heat load. This requires data centers to be equipped with more efficient liquid cooling systems and higher power supply densities.
Future: Bridge to the 1.6T Era
The large-scale deployment of 800G has accumulated valuable experience for the upcoming 1.6T (1600G) technology in 2025-2026. It validates the feasibility of new technologies such as PAM4 modulation and LPO/CPO, and forces the entire industry chain to upgrade in materials, testing, and operations.
800G is far more than just a speed standard. It is an inevitable product driven by AI and a culmination of collaborative innovation in multiple fields such as optical communication, semiconductors, and network protocols. For any organization wishing to maintain a leading position in future AI competitions, understanding and laying out an 800G data center network has shifted from an “optional” to a “mandatory” option. This silent ‘speed and passion’ is unfolding among the racks of data centers worldwide, and will ultimately determine how fast we can explore the boundaries of AI intelligence.
