Fujitsu Unveils AI Computing Broker to Optimize GPU Efficiency Amid Global Shortages

Fujitsu has introduced AI computing broker middleware designed to boost GPU efficiency and alleviate the challenges posed by the global GPU shortage. The new technology incorporates Fujitsu’s adaptive GPU allocator, a proprietary system that dynamically distributes GPU resources for real-time, high-efficiency AI processing. By combining this with various AI optimization methods, the solution addresses the growing demand for AI workloads while managing power consumption effectively.

---

Pilot Success and Industry Adoption

TRADOM Inc. will roll out solutions based on Fujitsu’s AI computing broker starting October 2024. Additionally, SAKURA internet Inc. has initiated a feasibility study to integrate this technology into its data center operations.

Earlier trials—conducted since May 2024 with companies like AWL Inc., Xtreme-D Inc., and Morgenrot Inc.—showed significant improvements. The trials revealed up to 2.25x better GPU processing performance and a substantial increase in the number of AI tasks that could be run simultaneously across multiple servers and cloud environments.

The AI computing broker will be available to customers in Japan starting October 22, 2024, followed by a global rollout.

---

Tackling AI-Driven GPU Demand and Energy Challenges

The rapid growth of AI technologies, including generative AI, has driven an unprecedented need for GPUs, which are more suitable for AI workloads than CPUs. Global demand for AI hardware is projected to grow 20 times by 2030, leading to increased power consumption. Current estimates suggest that data centers could account for 10% of the world’s electricity usage by 2030.

To address this issue, Fujitsu developed adaptive GPU allocation technology in November 2023. This technology optimizes CPU and GPU use by prioritizing processes with higher efficiency, even during active GPU workloads. This real-time optimization aims to increase computational output while reducing energy consumption.

---

AI Computing Broker: Key Features and Performance

The AI computing broker middleware integrates GPU allocation with advanced AI optimization techniques. It offers dynamic resource allocation at the GPU level rather than assigning resources on a per-job basis, resulting in higher availability and performance. The middleware’s memory management system also allows users to run multiple AI processes simultaneously without worrying about memory constraints, supporting workloads up to five times the physical GPU capacity (up to 150GB).

In trials, the technology improved GPU processing times by 2.25x compared to conventional methods. It also enabled parallel AI model training and testing, reducing overall execution time by nearly 10%.

---

Industry Endorsements and Use Cases

• TRADOM Inc.:
  “Our pilot with Fujitsu’s AI computing broker streamlined GPU allocation, significantly accelerating the development of our AI models for foreign exchange risk management,” said Junichi Kayamoto, Chief Data Science Officer at TRADOM. “This collaboration is essential for expanding our solutions and driving innovation in the FinTech sector.”
• SAKURA internet Inc.:
  Ken Wakishita, Senior Director of SAKURA Internet Research Center, noted: “The AI computing broker improved GPU efficiency within our cloud business, enabling broader access to GPU resources. We look forward to integrating this technology to meet growing market demand.”
• AWL Inc.:
  Hiroshi Fujimura, R&D General Manager at AWL, commented: “Optimizing GPU costs for parallel model training is critical for our AI-powered retail solutions. Fujitsu’s broker technology improved GPU utilization, helping us meet client demands efficiently.”
• Xtreme-D Inc.:
  Naoki Shibata, CEO and CTO, shared: “Our cloud platform Raplase caters to AI and HPC clients who need to optimize performance. Fujitsu’s AI computing broker addresses the challenge of efficient GPU utilization, and we are working closely with them to integrate it into our services.”
• Morgenrot Inc.:
  Masamichi Nakamura, COO, and Hisashi Ito, CTO, highlighted: “Our decentralized approach to cloud computing benefits from Fujitsu’s technology. During trials, we achieved faster execution by sharing GPU resources between jobs, unlocking new possibilities for simultaneous AI workloads. We look forward to further collaboration with Fujitsu.”

---

Future Expansion and Application

Fujitsu plans to extend the AI computing broker’s functionality to multiple GPUs across servers, anticipating broader use in large-scale computing environments. As part of its ongoing efforts, Fujitsu aims to tackle challenges related to GPU availability and energy efficiency, enabling sustainable AI development and boosting productivity for businesses.

For more information, visit www.fujitsu.com.