Report on "Research Strategy Briefing Session" held in June 2024 Creating new values by fusing technological areas around AI
Technology News | 2024-09-27
Fujitsu held the Research Strategy Briefing Session for the press, investors and industry analysts on June 4, 2024. Our R & D activities focus on five key technologies: AI, Computing, Data & Security, Networking, and Converging Technology. At this briefing session, we presented a new research strategy, "Creating new value by fusing AI and other key technology areas around AI."
This article introduces the contents of this briefing session. The three speakers are as follows.
- Research Strategy in the AI Area: Toshihiro Sonoda, VP, Head of Artificial Intelligence Laboratory, Fujitsu Research, Fujitsu Limited
- FUJITSU-MONAKA: Naoki Shinjo, SVP, Head of Advanced Technology Development Unit, Fujitsu Research, Fujitsu Limited
- Fujitsu's Research Strategy: Seishi Okamoto, EVP, Head of Fujitsu Research, Fujitsu Limited
Fujitsu AI research focuses on specialized models
Sonoda, VP, Head of Artificial Intelligence Laboratory, Fujitsu Research explained our research strategy in the AI field.
There are two trends in the area of generative AI technology, which has recently attracted much attention. One is the Large Language Model (LLM), typified by GPT. LLM supports video and audio in addition to language and is widely available in the cloud. The second is a specialized small-to-medium language model (SLM) to support enterprise operations. "Fujitsu's AI research will focus on specialized models that meet enterprise needs" Sonoda stressed.
There are several challenges with specialized enterprise models. Existing specialized models cannot handle the variety and volume of data companies have. They also face difficulties in rapidly generating business know-how and process-specific models, and in complying with corporate governance.
We provide a generative AI framework for enterprises that solves these issues of specialized models for corporate operations and eliminates security concerns.
The generative AI framework for enterprises consists of three technologies: Knowledge Graph Extended RAG, Amalgamation Technology, and Generative AI Audit Technology. These three technologies can be used alone, but Sonoda said that the three technologies work together to maximize value.
Knowledge Graph Extended RAG is our proprietary technology that processes a larger amount of data more efficiently than ordinary RAG (Retrieval Augmented Generation) and improves the accuracy of generated AI answers. RAG can be applied to large enterprise data sets by creating knowledge graphs to generate answers that look at the relationships between the data and the data as a whole. This technology has been ranked number one in the world in HotpotQA, a benchmark that measures the accuracy of complex question answering.
Amalgamation Technology automatically generates specialized generative AI without prompt engineering or fine-tuning. Automatically select the desired AI model from query (request to perform a task) and model characteristics. If a suitable AI model is not available, a new AI model is automatically generated. This technology is used to generate specialized AI that optimizes the placement of drivers in the transportation industry and that optimizes the placement of incident responders at the support desk.
Generative AI Audit Technology is a technology designed to ensure that generative AI systems comply with legal regulations and corporate policies. It works by creating a knowledge graph that encapsulates these regulations and policies, along with a specialized AI model that can assess compliance. When provided with input data, this technology can determine if any violations exist, providing both the judgment and the rationale behind it. By verifying the consistency between the generative AI's judgment and its rationale, this technology can accurately detect rule violations.
The Generative AI Framework for Enterprises is available starting in July.
FUJITSU-MONAKA software development ahead
Next, Shinjo, SVP, Head of Advanced Technology Development Unit, Fujitsu Research, gave a presentation on FUJITSU-MONAKA *, a processor for next-generation green data centers.
In recent years, the use of AI and big data has led to an increase in power consumption in data centers, which has a negative impact on the environment such as global warming. To address this issue, we are developing the FUJITSU-MONAKA, a processor for data centers that achieves both low power consumption and high performance.
FUJITSU-MONAKA is an Arm-based CPU utilizing 2-nanometer technology. It leverages Fujitsu's proprietary technologies, including low-voltage techniques, aiming for twice the processing performance and power efficiency compared to competitors. With its 144 cores and features like confidential computing, FUJITSU-MONAKA delivers the data processing power, reliability, and security features necessary for diverse workloads, including AI and HPC (High-Performance Computing), primarily targeting data centers.
Shinjo highlighted that Fujitsu is actively engaging in co-creation to enable the widespread adoption of FUJITSU-MONAKA across various fields. He emphasized their commitment to expanding the software ecosystem for the CPU, particularly focusing on software essential for AI and HPC, including machine learning, deep learning, data analytics, and data security. Furthermore, Fujitsu is collaborating with OSS community to develop Unified Acceleration technology, enabling the use of diverse AI accelerators with a single codebase. A recent example of this is the adaptation of oneDAL, previously reliant on Intel's numerical computation libraries, to function on Arm-based CPUs.
By proactively building a robust software ecosystem, fostering collaboration with the Arm ecosystem and OSS communities, and addressing software compatibility even before hardware shipments, Fujitsu ensures that FUJITSU-MONAKA is readily deployable in customer data centers upon its release.
Note: This achievement was made possible through funding from NEDO (New Energy and Industrial Technology Development Organization), a Japanese government-affiliated research and development agency.
Tackling Global Challenges Around AI
Finally, Okamoto, EVP, Head of Fujitsu Research, Fujitsu Limited, presented Fujitsu's overall research strategy.
Our R & D activities focus on five key technologies: AI, Computing, Data & Security, Networking, and Converging Technology. Okamoto explained his research strategy going forward: "Create new value by fusing these key technology areas with AI as the core."
In the area of technology convergence of AI and computing, we will address the global power problems associated with the development of AI. By 2030, 10% of all the earth's electricity will be consumed in data centers. In this context, we have developed technology to increase the efficiency of the use of GPUs, which make up the bulk of AI's computing resources. At present, even with the latest supercomputers equipped with various efficiency technologies, the GPU utilization rate remains at around 30%. Our AI Computing Broker technology enables 100% full GPU utilization. We estimate that the electricity saved by this technology is equivalent to the annual electricity consumption of approximately 24 million households in Japan.
In the area of AI and data & security technology fusion, we are addressing social risks posed by misinformation generated by generative AI and synthetic content. The World Economic Forum in 2024 identified ""AI-generated misinformation and disinformation"" as one of the greatest global risks. We are tackling this challenge through both rule-making and technological development. Regarding rule-making, we actively participate in discussions on international governance, including the ""G7 Hiroshima AI Process"" outlined at the 2023 G7 Hiroshima Summit and the ""AI Guidelines for Business"" being developed by the Japanese government. On the technological front, we are developing a system for counter-disinformation (Truth Verification Integrated Analysis System) to detect and identify misinformation and disinformation on the internet.
In the area of technology integration of AI and converging technology, we developed social digital twin technology that automatically formulates trade-on measures for the environment, society, and economy. We are currently conducting social demonstrations in the fields of mobility, energy environment, disaster prevention and crime prevention, and well-being. With our ""real-time 3D twin generation"" technology, you can create a 3D twin of a city in real time with a single monocular camera.
In the area of AI and quantum computing technology fusion, quantum computing's extraordinary computational power will contribute significantly to the advancement of AI. Our company boasts world-leading technology in quantum machine learning, including the fastest quantum CNN (Convolutional Neural Network) technology. We were also the first in the world to successfully develop "quantum autoencoder technology" for quantum noise reduction. Okamoto highlighted this area as a unique fusion domain where we possess top-tier expertise in both fields.