Fujitsu has introduced a new AI computing broker middleware aimed at improving GPU utilisation and addressing the global shortage of GPUs. The technology has shown impressive results in trials, significantly enhancing AI processing efficiency by dynamically allocating GPU resources across various platforms and AI applications.
Enhanced efficiency through dynamic GPU allocation
Fujitsu’s AI computing broker middleware is designed to optimise GPU computational efficiency, which is crucial for handling the growing demand for AI technology. Trials conducted by Fujitsu with partners like AWL, Inc. and Xtreme-D Inc. demonstrated up to a 2.25x increase in computational efficiency, significantly improving AI processes. This was achieved through the middleware’s ability to dynamically allocate GPU resources on a per-GPU basis, optimising memory management and resource distribution.
One standout feature of the new technology is its capability to handle up to 150GB of AI processing simultaneously, approximately five times the physical memory capacity of a standard GPU. This allows multiple AI processes to run without overloading the system’s resources, essential in today’s data-intensive AI landscape.
Industry applications and trial results
Fujitsu’s trials with various companies highlighted the middleware’s effectiveness in enhancing AI processing across cloud and on-premise environments. TRADOM Inc., for example, plans to implement the AI computing broker for its AI-powered foreign exchange risk management solutions from October 2024, based on positive trial outcomes. Similarly, SAKURA Internet Inc. has started a feasibility study to apply this technology in their data centres.
Morgenrot Inc. is also trialling the middleware, which has delivered remarkable results. By enabling parallel processing, it has reduced execution time by nearly 10%. By allowing the simultaneous execution of AI model training and shorter inference tasks, this technology is expected to revolutionise GPU usage, making it more efficient and cost-effective.
Tackling GPU shortages and power consumption
As global demand for GPUs continues to rise, driven by the rapid growth of AI and generative AI, Fujitsu’s AI computing broker middleware could play a crucial role in addressing GPU shortages. The company’s adaptive GPU allocator technology, first developed in November 2023, optimises the use of both CPUs and GPUs, ensuring that GPU resources are prioritised for high-efficiency processes.
Given the increasing power consumption of data centres, Fujitsu’s technology also aims to reduce the strain on resources. With data centres projected to consume 10% of the world’s electricity by 2030, solutions that improve efficiency, like the AI computing broker, will be critical in managing the rising costs and environmental impact of AI processing.
Fujitsu has made the AI computing broker available to customers in Japan as of 22 October 2024, with plans to expand globally. This technology will benefit AI service providers looking to reduce GPU costs and cloud service providers aiming to maximise resource utilisation.
Looking ahead
Fujitsu plans to continue expanding the scope of its AI computing broker middleware, anticipating future applications across larger computing environments and multiple GPUs. The company is committed to developing innovative technologies that address global challenges such as GPU shortages and rising energy consumption, all while boosting business productivity and creativity for its customers.
Industry experts, including Junichi Kayamoto of TRADOM Inc. and Ken Wakishita of SAKURA Internet, have expressed confidence in the technology’s ability to transform AI processing. These partnerships will help drive the adoption of Fujitsu’s AI computing broker middleware, offering businesses a more efficient and sustainable approach to GPU utilisation.
