Latest News

  • Research Highlights

  • Publish Date:2026-07-27
HHRI and NYCU Unveil Silicon Photonic Breakthrough for Next-Generation AI Data Centers
A concept illustration of ultra-high-speed optical data transmission. The proof-of-concept silicon photonic transmitter achieved 34.132 Tbit/s through a single optical fiber—equivalent to transferring thousands of high-definition movies in just one second.
A concept illustration of ultra-high-speed optical data transmission. The proof-of-concept silicon photonic transmitter achieved 34.132 Tbit/s through a single optical fiber—equivalent to transferring thousands of high-definition movies in just one second.
 
Edited by Chance Lai
______
As artificial intelligence models become larger and more computationally demanding, the next bottleneck in AI is no longer processing power alone—it is how quickly massive amounts of data can move between chips, servers and entire data centers.

Researchers at National Yang Ming Chiao Tung University (NYCU) and the Hon Hai Research Institute (HHRI) have developed a high-capacity silicon photonic transmitter that combines a quantum-dot comb laser, wavelength-division multiplexing and multicore fiber technologies to increase optical data transmission. The proof-of-concept system achieved an aggregate transmission capacity of 34.132 Tbit/s through a single optical fiber, providing a promising architecture for future AI data centers and co-packaged optics (CPO). The study was published in Optics Express.

Replacing Electrons With Light

The rapid rise of large language models (LLMs) and generative AI has transformed the architecture of modern AI data centers, where thousands—or even tens of thousands—of GPUs work together to train and deploy increasingly sophisticated models. As computing power continues to scale, however, the volume of data exchanged between processors grows even faster. Moving data efficiently has therefore become one of the industry’s greatest engineering challenges.

To overcome this bottleneck, the semiconductor industry has increasingly turned to silicon photonics, which uses light instead of electrical signals to transmit data at much higher speeds. Unlike conventional copper interconnects, which consume more power, generate more heat and suffer greater signal loss over longer distances, optical communication can carry substantially more information with significantly higher energy efficiency. As a result, silicon photonics is widely regarded as one of the key enabling technologies for next-generation AI infrastructure.

The joint NYCU–HHRI research team sought to increase transmission capacity further while reducing the number of optical components required, creating a simpler, more compact and energy-efficient optical transmission architecture.

One Laser, One Fiber, 34.132 Tbit/s

One of the team’s key innovations is the use of an ultra-broadband quantum-dot comb laser. Conventional optical communication systems typically require multiple independent lasers to generate different wavelengths. In contrast, the new transmitter uses a single laser to stably produce 23 distinct optical wavelengths, greatly simplifying the optical architecture.

The system can be likened to a single light source generating 23 independent high-speed optical channels, each capable of carrying data simultaneously without interfering with the others. By replacing multiple lasers with a single light source, the design reduces system complexity while easing challenges related to packaging, thermal management and device control.

Each wavelength employs PAM4 (Pulse Amplitude Modulation 4-level), enabling transmission at 212 Gbit/s while carrying twice as much information as conventional binary modulation within the same bandwidth. The researchers further expanded system capacity by incorporating a seven-core multicore optical fiber. Unlike conventional optical fibers, which contain only a single transmission path, multicore fibers integrate multiple independent cores within one fiber, allowing several data streams to travel simultaneously.



By combining 23 wavelength channels with seven spatial cores, the transmitter simultaneously employs wavelength-division multiplexing (WDM) and space-division multiplexing (SDM), dramatically increasing overall throughput. Together, the system achieved an aggregate transmission capacity of 34.132 Tbit/s through a single optical fiber. To validate the architecture, the team successfully transmitted data over 2 kilometers of optical fiber while maintaining excellent signal integrity.

The researchers emphasize that the current work represents a proof of concept demonstrating the feasibility of the integrated architecture. Before commercialization, additional advances in device packaging, long-term reliability and scalable manufacturing will still be required.

Toward the Next Generation of AI Infrastructure

Beyond boosting transmission capacity, the researchers also designed specialized high-frequency electrodes within the silicon photonic chip to suppress electromagnetic interference generated during simultaneous multi-channel operation, further improving signal integrity.

The technology could contribute to the development of co-packaged optics (CPO), an emerging architecture that places optical components much closer to AI processors and network switches. By shortening electrical transmission distances, CPO can significantly reduce power consumption and latency while overcoming one of the major performance bottlenecks in future AI computing systems.

As AI models continue to grow, demand for faster and more energy-efficient optical interconnects will only increase. By combining advanced silicon photonics with innovative optical transmission architectures, the collaboration between NYCU and HHRI demonstrates Taiwan’s growing leadership in silicon photonics and high-speed optical communications, while laying the groundwork for the next generation of AI data center infrastructure.

The research was led by Hao-Chung Kuo, director of the Semiconductor Research Institute at HHRI, and Chi-Wai Chow, distinguished professor at NYCU, together with researchers from both institutions. The project was supported by Taiwan’s National Science and Technology Council (NSTC) and the Industrial Technology Research Institute (ITRI), with additional research collaboration from the team led by Chair Professor Wood-Hi Cheng at National Chung Hsing University.

A schematic illustration of co-packaged optics (CPO), where optical engines are integrated closer to AI processors to shorten electrical interconnect distances, improving bandwidth, energy efficiency and signal integrity.A schematic illustration of co-packaged optics (CPO), where optical engines are integrated closer to AI processors to shorten electrical interconnect distances, improving bandwidth, energy efficiency and signal integrity.
資料來源/鴻海研究院
文圖/國際宣傳辦公室 


當AI快速發展,全球資料中心持續擴建,運算能力早已不再只是比拚GPU有多快,如何讓龐大的資料在晶片、伺服器與機櫃之間快速流動,正成為下一個關鍵挑戰。

國立陽明交通大學攜手鴻海研究院,成功開發新一代高容量矽光子發射器(Silicon Photonic Transmitter),透過單一量子點梳狀雷射、多波長光訊號與多核心光纖整合,大幅提升資料傳輸效率,為未來AI資料中心與共同封裝光學(Co-Packaged Optics, CPO)提供新的技術方向。研究成果已發表於國際光學期刊《Optics Express》。

從「電子」改成「光」 讓資料跑得更快

近年大型語言模型(LLM)與生成式AI快速發展,一座AI資料中心往往需要數千甚至數萬顆GPU共同運算。然而,GPU愈多,需要交換的資料量也同步暴增。如果資料無法即時送達,再快的運算晶片也必須等待資料傳輸完成,導致整體效能受到限制。因此,全球科技產業近年積極投入矽光子(Silicon Photonics)技術,希望利用光訊號取代部分電訊號,降低高速傳輸造成的能耗、散熱與訊號損失。

傳統晶片之間主要透過銅線傳送電子訊號,就像高速公路上的車流,距離愈長、速度愈快,越容易塞車、耗能,也更容易發熱。矽光子技術則改以光來傳遞資料。由於光可以承載更多資訊,且長距離傳輸時能維持較低的能量損耗,因此被視為下一代AI資料中心的重要核心技術。這次陽明交大與鴻海研究院的研究,便希望利用更少的光源與光纖,傳送更多資料,同時降低系統複雜度。

一顆雷射、一根光纖 打造34.132 Tbit/s高速傳輸

研究團隊最大的突破之一,是採用超寬頻量子點梳狀雷射(Quantum-dot Comb Laser)。一般光學系統若需要多個波長,通常必須配置多顆獨立雷射;此次研究則只利用一顆雷射,便能穩定產生23個不同波長的光訊號。可以把它想像成:一盞燈,同時分出23條不同顏色、彼此互不干擾的高速光通道。如此一來,不僅能減少雷射數量,也能降低光學系統在封裝、散熱與控制上的複雜度。
 



此外,每個波長皆採用PAM4高速調變技術,每秒可傳輸212 Gbit資料,相較於傳統傳輸方式,可在相同頻寬下承載更多資訊。除了增加波長數量外,研究團隊另一項創新,是採用7核心多核心光纖(Multi-core Fiber)。一般光纖只有一條傳輸路徑,而多核心光纖則像是在同一條光纖內建置多條獨立車道,可讓多組資料同時傳輸。當23個波長再搭配7個光纖核心後,整體系統可同時利用**波長多工(WDM)與空間多工(SDM)**兩種技術,大幅提升傳輸容量。最終,單根光纖總傳輸容量達到34.132 Tbit/s(每秒34.132兆位元)。

研究團隊也完成2公里光纖傳輸測試,證實高速傳輸後仍能維持良好的訊號品質。不過,研究人員也指出,這項成果目前仍屬概念驗證(Proof of Concept),主要目的在於驗證整合架構的可行性,距離商業化應用仍需克服封裝、可靠度及量產等挑戰。

為下一代AI資料中心奠定基礎

除了高速傳輸外,研究團隊也在矽光子晶片中導入特殊高頻電極設計,降低多通道同時運作時的電磁干擾,提升整體訊號穩定性。研究成果未來可望應用於共同封裝光學(CPO)技術,將光學元件更靠近AI晶片與交換器,降低高速電訊號傳輸造成的能耗與延遲。

隨著AI模型持續擴大,資料中心對高速互連技術的需求也將快速增加。此次陽明交大與鴻海研究院共同完成的研究,不僅展現台灣在矽光子與高速光通訊領域的研發實力,也為下一代AI運算基礎建設提供新的技術方向。

本研究由鴻海研究院半導體所所長郭浩中、小組長洪瑜亨、研究員張雲翰,以及陽明交大特聘教授鄒志偉共同完成,並獲國家科學及技術委員會、工業技術研究院及中興大學講座教授鄭木海研究團隊支持。

Related Image(s):