In a notable advancement that may redefine technology in data centers, IBM has recently unveiled an innovative polymer waveguide (PWG) technologyThis breakthrough pertains to co-packaged optics, a field critical for optimizing data transmission speeds within complex computing environments.

This new optical technology promises to significantly change how data centers train and execute generative AI modelsEssentially, it integrates fiber optic power directly onto the chip, facilitating connections at light speed—expediting communications among the components within data centers.

Mukesh Khare, IBM's General Manager of Semiconductor research, highlighted the disparity in the growth rate of semiconductor manufacturing versus communication speeds between chips

While the telecom sector has made substantial progress in developing faster chips, the corresponding communication technologies have lagged behind.

He noted, "Fundamentally, more basic chips still communicate through electrical signals using copper wiresWe know that our best communication technology is fiber optics, which is why they are used for long-distance communication in other fields."

Despite co-packaged optics being in existence for some time, IBM's PWG technology fundamentally enhances this by enabling chip manufacturers to add over six times more fiber optics to silicon photonic chipsEach fiber is about three times the width of a human hair, and they can span lengths from a few centimeters to hundreds of meters, transmitting data at several terabits per second.

This presents an extraordinary opportunity: IBM claims that the implemented technology can deliver communication bandwidth between chips that is 80 times faster than current electrical communication methods, while also achieving a reduction in energy usage by over five times.

Moreover, it holds the potential to increase the training speed of large language models (LLMs) by fivefold, reducing the time needed for training standard LLMs from three months down to just three weeks, thus enhancing performance through the employment of larger models and more GPUs.

Not only does this technology enable faster communication between GPUs and accelerators, but it could also redefine how the computing industry transmits high-bandwidth data across circuit boards and servers.

Khare expressed enthusiasm about capitalizing on the power of optics to accelerate Gen AI and numerous other applications.

When questioned about the commercialization timeline of this technology, Khare stated that IBM's research department is ready for deployment.

The evolution of electronic chips has faced extensive challenges, particularly as voices of "the end of Moore's Law" continue to rise

• China's Chip Industry Sees Progress
• The Logic Behind Global Commodity Pricing
• Guoxuan Deal Adds $1.2 Billion to Volkswagen Profit
• Dow Jones Posts Nine-Day Losing Streak
• Argentina's GDP Up Quarter-on-Quarter, Down Year-on-Year

Silicon-based electronic chips run into physical limitations around 7 nanometers, where issues like electrical surges and electron breakdown become predominant.

In contrast, photonic chips present a new paradigm, enabling breakthroughs in power consumption and memory capabilitiesThey hold the promise of unlocking unprecedented applications across various industries.

Currently, intense competition rages among top research institutions globally over photonic chip developmentA noteworthy instance is the research team from Tsinghua University, which, in April 2023, proposed an innovative distributed breadth-intelligent optical computing architecture, resulting in the “Taiji” photonic chip, which boasts energy efficiency surpassing existing intelligent chips by several orders of magnitude.

The workings of photonic chips diverge markedly from traditional electronic chips, which rely on electronic transistors and conductive wires

Instead, photonic chips utilize photonic transistors and light waveguidesWaveguides facilitate the propagation of optical signals, akin to common optical fibers.

Despite their promise, pure photonic chip research is still largely experimental, and most existing photonic chips require electricity for controlNonetheless, Tsinghua University’s “Taiji II” chip has demonstrated potential for online training via optical neural networks, achieving high-speed data processing without GPUsThis presents a hopeful avenue for the practical application of photonic chips.

The journey to commercialize photonic chips is riddled with technical and cost challenges