Japanese telecommunications firm NTT has unveiled a range of innovative research projects aimed at enhancing artificial intelligence (AI) and improving energy efficiency in data centers.
During a recent press conference in San Francisco, NTT researchers introduced a newly developed large language model (LLM) that can analyze graphical elements within documents. They also announced the initiation of a groundbreaking field of research labeled “the physics of intelligence,” which focuses on developing sustainable and trustworthy AI, in collaboration with Harvard University.
At the event, NTT showcased its all-photonic network, designed for distributed data centers, achieving significant advancements in low-latency communication. Yosuke Aragane, NTT’s vice president of the IOWN Development Office, emphasized that relocating large data centers to suburbia could lower costs and enhance energy efficiency, particularly through NTT’s rapid fiber connections.
NTT, with a workforce exceeding 330,000 and annual revenues of $97 billion, invests over $3.6 billion each year in research and development. The company established its R&D division in Silicon Valley five years ago, showcasing its progress at the Upgrade 2024 event in San Francisco.
“Our mission is to elevate your concept of normalcy to the next level,” stated Kazu Gomi, president and CEO of NTT Research.
Low-Latency Networks in the U.S. and U.K.
Aragane highlighted that NTT's IOWN All-photonics network (APN) demonstration achieved remarkably low communication delays between connected data centers, making it essential for AI analysis and financial services. He noted that urban areas face challenges like high costs, land shortages, and expensive electricity. NTT is exploring the feasibility of distributing data centers to the suburbs, connecting them with fiber-optic cables that support data transfer rates of up to 400 gigabits per second.
In the U.K., tests demonstrated that data centers located 100 kilometers apart experienced less than one millisecond of network delay with APN connections. This network performance significantly minimized delay variation compared to traditional networks, bringing geographically dispersed data centers closer in functionality to a single center.
The U.K. has connected data centers north and east of London using NTT's Innovative Optical Wireless Network (IOWN) APN, achieving a round-trip delay of less than one millisecond. Similar results were observed in Northern Virginia, U.S.
The initiative aims to transform distant IT infrastructures into a unified data center experience. Local constraints such as carbon dioxide emission regulations and space limitations are driving operators to seek suburban alternatives. However, latency remains a challenge with geographically dispersed centers. NTT's photonic links are being utilized to address these issues.
Advanced Data Center Demonstrations
In separate trials, NTT and NTT DATA successfully connected data centers in the U.K. and the U.S. using APN technology. Tests revealed that the U.K. data centers operated with less than one millisecond of latency and remarkably low delay variation. For comparison, conventional networks typically exhibit latency exceeding 2,000 microseconds.
The APN system meets the stringent latency requirements for current and emerging applications, including real-time AI analysis for industrial IoT, smart energy management, and natural disaster response. In the financial sector, NTT DATA is running demonstrations where low latency is crucial for transactions and remittances.
Aragane remarked, “Demand for data centers is increasing, yet land scarcity and electricity availability hinder new facility development. We aim to create more energy-efficient data centers to tackle these obstacles.”
Innovations in Visual Comprehension
NTT also introduced a breakthrough in visual machine reading comprehension, enabling LLMs to interpret graphical elements within documents, including charts and diagrams. Developed in collaboration with Jun Suzuki from Tohoku University, this technology has been applied to NTT's lightweight LLM, tsuzumi.
When benchmarked against major models, NTT's LLM surpasses open-source options like LLaVA and OpenAI’s GPT-3.5 and GPT-4 across various document understanding tasks. Kyosuke Nishida, a senior researcher at NTT, acknowledged the growing capability of LLMs while also addressing ongoing challenges in processing multimodal information.
First unveiled in November 2023, tsuzumi is available in two variants: an ultra-lightweight version with 600 million parameters and a lightweight version with seven billion parameters. Its compact size drastically reduces energy consumption and training costs, making it a sustainable choice for enterprises.
Potential applications of tsuzumi include automation in call centers, digital record-keeping, and tasks in software engineering. The model currently supports over 20 languages, including English and Japanese, and is undergoing commercial trials with more than 500 global companies.
Partnership with Harvard's Center for Brain Science
NTT Research has committed a significant donation to establish the Harvard University Center for Brain Science (CBS)-NTT Fellowship Program, fostering post-doctoral research in the emerging field of the physics of intelligence. This renewable two-year gift could ultimately exceed $1.7 million, supporting innovative research at the nexus of computer science, neuroscience, and psychology.
NTT's collaboration with Harvard CBS has previously yielded valuable insights, including addressing AI bias using cognitive science principles. Recent publications have explored the science behind generative AI and its applicability to neuroscience.
“Supporting the Harvard CBS aligns with our vision of leveraging AI to address pressing issues like computational fairness and sustainability,” said Gomi.
A Unique Position in Silicon Valley
With a substantial office in Sunnyvale, California, NTT Research distinguishes itself through its commitment to foundational research. Over the past five years, NTT Research has published over 450 academic papers, receiving multiple awards for excellence across various scientific disciplines.
Gomi indicated that ongoing research focuses on developing photonic integrated circuits and studying brain function to better understand computational processes. Additionally, NTT aims to advance quantum-resistant encryption technology to prepare for future quantum computing challenges.
As NTT pushes forward, it envisions creating “digital twins” of biological systems, such as the heart, to simulate drug responses for personalized medical care.
Through these strategic initiatives, NTT is not only advancing technology but also enhancing the future of AI and data center operations.
Ranking
