Race to 448 Gbps

The relentless growth in data center and AI workloads is accelerating the need for faster, more efficient interconnect technologies. As applications like large language models and distributed training infrastructures push the limits of bandwidth, 400/800G Ethernet is quickly becoming a bottleneck. To support next-generation performance at 1.6T and 3.2T system levels, engineers must enable reliable 448 Gbps-per-lane signaling through a combination of advanced modulation, signal integrity strategies, and cutting-edge test solutions.

This white paper investigates the technical and practical challenges involved in achieving 448 Gbps transmission. The discussion begins with Shannon’s Law, which sets the theoretical maximum data rate of a system based on available bandwidth and signal-to-noise ratio (SNR). While this law defines the upper limit, real-world viability hinges on Symbol Error Rate (SER), which is highly sensitive to both the chosen modulation format and actual channel conditions. Higher-order PAM formats such as PAM6 and PAM8 can significantly boost throughput—by 1.29× and 1.5× over PAM4 respectively—at the same symbol rate. However, this efficiency comes at the cost of increased SNR requirements and signal degradation from jitter, ISI, and noise.

A detailed comparison of NRZ, PAM3, PAM4, PAM6, and PAM8 modulation schemes illustrates the trade-offs in terms of bits per symbol, bandwidth requirements, and error performance. While PAM4 is widely deployed and offers a solid balance of efficiency and complexity, scaling to 448 Gbps requires more aggressive strategies. For instance, achieving 448 Gbps with PAM4 necessitates symbol rates as high as 224 Gbaud—pushing test equipment to its limits. Alternatively, PAM6 and PAM8 can achieve 448 Gbps at lower symbol rates (174 Gbaud and 150 Gbaud, respectively), easing bandwidth constraints but requiring pristine signal quality.

The paper further delves into the signal integrity and test challenges that accompany these higher-order modulation schemes. As the number of signal levels increases, eye diagrams become more compressed, and equalization demands grow. To ensure reliable transmission, advanced equalization techniques—such as CTLE, FFE, and DFE—must be employed in both transmitter and receiver design. Visual comparisons of eye diagrams for PAM4, PAM6, and PAM8 highlight the narrowing eye openings and growing susceptibility to noise.

Keysight’s role in advancing high-speed interconnect test strategies is showcased through its industry-leading Arbitrary Waveform Generator (AWG) solutions, including the M8199B platform. These systems are purpose-built for PAM4/6/8 generation and analysis at symbol rates beyond 200 Gbaud. The paper presents benchmark results demonstrating over 120 GHz of usable bandwidth, SNR above 20.5 dB at 224 Gbaud, and sample rates up to 512 GSa/s. Using Keysight’s Frequency Domain Interleaving Unit (M8159A), researchers can extend signal generation bandwidth even further to support hero experiments at ≥ 240 Gbaud, which are essential for pushing interconnect technologies into the 3.2T era.

Validation techniques are also explored in both electrical and optical domains. The paper outlines recommended test setups for electrical signal generation and analysis—including high-bandwidth AWGs and ultra-low-jitter oscilloscopes—as well as optical validation tools capable of supporting 200–400 Gbps per-lane testing. This comprehensive test ecosystem ensures that developers can analyze link behavior, validate interoperability, and fine-tune systems for peak performance under real-world conditions.

Ultimately, the path to 448 Gbps and beyond is not simply a matter of faster signals. It demands a complete rethinking of modulation efficiency, signal quality, instrumentation, and test methodology. By advancing both signal generation and measurement capabilities, Keysight enables engineers to navigate this complexity and bring next-generation interconnect solutions to life. These breakthroughs are critical for AI infrastructure, cloud computing, and telecommunications systems that must scale quickly and efficiently without sacrificing reliability or performance.