- Fujitsu’s groundbreaking technology addresses the global GPU shortage by optimizing CPU and GPU resource allocation in real time.
- This innovation is valuable for accelerating AI development and high-performance computing, benefiting industries and researchers.
- Fujitsu’s future plans include applying the technology to AI platforms, quantum computing, and more, advancing accessibility to high-performance computing resources.
Fujitsu, a global leader in technology and innovation, has revealed a game-changing technological improvement that promises to transform CPU and GPU utilization. This cutting-edge technology, the first of its type in the world, dynamically distributes computing resources in real time, prioritizing tasks with great execution efficiency. This invention comes at a critical juncture in the world’s GPU shortage, which has been compounded by rising demand for generative AI, deep learning, and other GPU-intensive applications.
Fujitsu’s novel approach promises to maximize existing computer resources while also providing a lifeline to enterprises and researchers who rely on high-performance computing (HPC). In addition, the company pioneered real-time parallel processing technology, which allows for the simultaneous execution of numerous applications on HPC systems without the need to wait for ongoing processes to complete. This discovery has the potential to dramatically improve the efficiency of large-scale computations, particularly in disciplines such as digital twin modeling and generative AI applications.
Managing the Global GPU Shortage
Global GPU shortages have posed a significant problem to enterprises and researchers who rely on these processing units for intensive applications such as generative AI and deep learning. Fujitsu’s solution addresses this issue head-on by distributing CPU and GPU resources efficiently in real-time, ensuring that high-priority programs have the processing power they require, even in a multi-program environment.
CPU and GPU Utilization During Program Execution
While numerous programs run concurrently, Fujitsu’s technology distinguishes between those that require GPU acceleration and those that can be executed by a CPU. It accomplishes this by estimating the pace of acceleration and allocating GPUs in real-time to high-priority program operations. Here’s a simple illustration of how it works:
- The customer intends to run three apps on one CPU and two GPUs.
- Initially, GPUs are assigned to programs 1 and 2 based on availability.
- In response to the request of program 3, GPU allocation is switched from program 1 to program 3 for performance evaluation.
- The system monitors the amount of processing acceleration on the GPU and discovers that dedicating the GPU to program 3 reduces total processing time.
As a result, the GPU is assigned to program 3, while the CPU is assigned to program 1. Following the completion of program 2, the GPU is transferred to program 1, maximizing computational resources and minimizing processing time.
This game-changing technique is especially useful for speeding up processes such as training AI models and analyzing graph AI data, ultimately boosting the development of AI apps and improved picture recognition.
Real-time Execution Switching in HPC Systems
In addition to solving the GPU scarcity, Fujitsu has created a game-changing technology that allows for real-time switching between several programs on HPC systems. Unlike old approaches, which require waiting for one program to finish before switching, this innovation enables for immediate program execution with no delay.
HPC systems have traditionally used unicast communication, which switched program execution sequentially, resulting in timing fluctuations and obstacles in attaining real-time program execution. Fujitsu’s approach makes use of broadcast transmission to enable simultaneous communication and program execution switching. In a 256-node HPC environment, this cuts the time between program processing changes from several seconds to just 100 milliseconds. Users can choose the appropriate connection channel based on application requirements and network conditions, assuring peak performance.
This technology provides the path for HPC-like computational resources to be used to swiftly execute applications requiring real-time performance, such as digital twin modeling, generative AI, and materials and drug development.
Plans and Consequences
Fujitsu plans to include CPU/GPU resource optimization technology in its Fujitsu Kozuchi (code name) – AI Platform. This platform allows users to quickly test complex AI technologies, hence speeding up AI research and development.
Fujitsu’s 40-qubit quantum computer simulator for collaborative computing will benefit from HPC optimization technologies, improving its capabilities and potential applications.
Fujitsu is also looking for broader uses for its technology. Fujitsu Computing as a Service HPC provides consumers with low-cost, high-performance computing capabilities for simulation, artificial intelligence, and combinatorial optimization challenges. Furthermore, Fujitsu is investigating the use of this technology to the Composable Disaggregated Infrastructure (CDI) architecture, which would allow customers to quickly adjust hardware configurations among servers. These initiatives are consistent with Fujitsu’s objective to provide a more accessible, sustainable, and high-performance computing ecosystem.