Taiwan researchers develop 100Gbps chip transceiver to support next-generation AI data centers

A research team led by Associate Professor Pen-Jui Peng (center right) from the Department of Electrical Engineering at National Tsing Hua University (NTHU) has developed a Pulse Amplitude Modulation-4 (PAM-4) transceiver with support from Taiwan's National Science and Technology Council (NSTC). The technology could replace conventional pluggable optical transceiver architectures and enable higher data transmission speeds with lower power consumption.

Pictured from left to right (starting from the second person): Ming-Wei Lin, Associate Research Fellow, Taiwan Semiconductor Research Institute (TSRI); Yi-Chun Liu, Associate Professor, Institute of Electronics Engineering, National Tsing Hua University; Kuang-Hung Lin, Secretary-General, National Science and Technology Council (NSTC); Pen-Jui Peng, Associate Professor, Department of Electrical Engineering, National Tsing Hua University; and Ping-Hsuan Hsieh, Associate Professor, Department of Electrical Engineering, National Tsing Hua University.

A Taiwanese research team has developed a 100-gigabit-per-second (Gbps) pulse amplitude modulation (PAM-4) transceiver and co-packaged optics (CPO) module, aiming to improve data transmission efficiency and reduce power consumption in artificial intelligence (AI) data centers.

The project was supported by Taiwan's National Science and Technology Council (NSTC) under its Key Emerging Chip Design Research and Development Program. The team includes researchers from National Tsing Hua University and the National Applied Research Laboratories' Taiwan Semiconductor Research Institute.

As demand for AI computing continues to grow, data centers are facing increasing pressure to transfer larger volumes of data at higher speeds while limiting energy consumption. Industry experts see optical interconnect technologies as a key solution to overcoming bandwidth and power constraints in future computing systems.

The newly developed PAM-4 transceiver achieves a single-lane transmission speed of 100Gbps. PAM-4 technology increases the amount of information carried by each signal by using four amplitude levels instead of the two levels employed in conventional signaling methods.

To address the complexity of decoding multi-level signals, the researchers designed a receiver architecture based on low-resolution analog-to-digital converters, enabling high-speed data sampling and signal recovery while reducing overall power consumption.

According to the research team, the transceiver can deliver performance comparable to solutions developed by major international high-speed chip manufacturers, despite being fabricated using a 28-nanometer process technology rather than more advanced sub-7nm nodes. The chip also incorporates self-calibration functions designed to compensate for temperature variations and manufacturing inconsistencies, improving reliability for large-scale production.

In addition to the transceiver, the team integrated the chip with silicon photonics components through advanced heterogeneous packaging technology to create a co-packaged optics module.

Unlike traditional pluggable optical modules, CPO technology places optical and electronic components closer together, shortening transmission distances and reducing energy losses. The approach is widely viewed as a promising architecture for future AI servers and high-performance computing systems.

The researchers also developed a silicon-based micro-ring modulator capable of 100Gbps electro-optical conversion and integrated a high-speed photodetector with bandwidth exceeding 50GHz.

A 112Gb/s co-packaged optics (CPO) system board and a 112Gb/s transceiver evaluation board.

Optical interconnect technologies have attracted growing attention from the semiconductor industry as AI workloads drive demand for faster and more energy-efficient communication between processors, accelerators and memory systems.

The project has resulted in six U.S. invention patents and eight Taiwan invention patents. The team said the technology has also led to multiple industry-academia collaboration projects and could be further commercialized for future applications in data center and high-performance computing infrastructure.

As global technology companies invest heavily in AI infrastructure, advances in optical communications and silicon photonics are expected to play an increasingly important role in improving computing efficiency and reducing energy consumption.

Source: NSTC News