Fujitsu has presented an innovative artificial intelligence (AI) middleware technology that optimizes the performance of GPUs, especially in a context of growing global scarcity of these resources. This tool, based on intelligent and dynamic GPU allocation developed by Fujitsu, allows adaptive distribution of GPUs in real time to maximize their efficient use and accelerate AI processing through advanced optimization techniques.
After successfully completing a series of pilot tests, TRADOM Inc. will begin implementing solutions based on this technology starting in October 2024. Similarly, SAKURA internet Inc. has initiated a feasibility study to evaluate the incorporation of this technology. technology in your data center operations. Additionally, Fujitsu has since May 2024 conducted performance testing with companies such as AWL, Inc., Xtreme-D Inc., and Morgenrot Inc., observing notable improvements in their operations, including a 2.25 times increase in computational efficiency of certain AI processes, as well as a significant increase in the number of processes managed simultaneously in server and cloud infrastructures.
Commercial availability of this technology in Japan is scheduled for October 22, 2024, and will be extended to other global markets in later stages. Fujitsu aims to continue supporting both end users and cloud service providers looking to reduce their GPU costs and increase their operational capabilities in AI.
Addressing the growing demand for AI and the energy challenge
With the exponential growth of AI applications, especially generative AI, the need for GPUs has grown considerably, and demand is expected to increase approximately 20-fold by 2030. This has created additional concern regarding energy consumption in centers of data, whose demand could represent 10% of global electricity by 2030.
The middleware dynamically allocates GPU resources on a per-GPU basis, optimizing resource allocation and memory management across various AI platforms and applications.
To address this challenge, Fujitsu developed its adaptive GPU allocation technology in November 2023, capable of optimizing the use of CPU and GPU resources in real time, prioritizing processes with high execution efficiency. This technology has already been tested on multiple platforms, showing its ability to maximize the use of infrastructure.
Features of the new AI intermediation middleware
The new middleware acts as an intelligent intermediary, integrating adaptive GPU allocation with advanced processing optimization techniques, allowing the allocation of GPUs between multiple AI processes to be automatically managed and optimized.
Instead of allocating resources at the task level, this brokering technology distributes resources at the individual GPU level, enabling higher availability and better management of GPU memory. This makes it possible for companies to run numerous AI processes simultaneously, even under conditions of high resource demand.
Testing demonstrated up to a 2.25x increase in GPU processing power and the ability to handle up to 150GB of data in AI processes, far exceeding the physical capacity of the GPU.
