Fujitsu introduces dynamic resource allocator for AI servers and HPC systems

Forward-looking: Fujitsu's business has traditionally focused connected trading computing products, mainframe servers, and HPC devices. Now, nan institution is aiming to leverage its "computational optimization expertise" to connection innovative, software-based solutions to reside nan increasing request for GPUs successful cutting-edge AI applications.

After introducing its "computing broker" solution successful 2023, Fujitsu has now confirmed that nan merchandise is yet disposable for acquisition successful Japan and different markets worldwide. The Kawasaki-based corp intends to execute done package what has traditionally been handled by hardware, which is expected to importantly amended assets optimization and GPU utilization.

The caller exertion is presented arsenic middleware designed to dynamically allocate resources connected a per-GPU basis, optimizing utilization and precocious representation guidance crossed aggregate platforms and AI applications. The computing agent allocates CPU and GPU computing resources successful real-time, prioritizing processes pinch higher execution efficiency. It tin besides reallocate processes moreover while they are already moving connected a GPU.

In pre-release testing, Fujitsu reportedly achieved up to a 2.25x betterment successful GPU processing performance. The exertion besides offers awesome representation guidance capabilities, arsenic it is designed to grip AI workloads of up to 150GB – astir 5 times nan beingness representation capacity of nan tested GPUs.

Fujitsu initially stated that developers needed to usage its proprietary model to afloat utilization nan capabilities of nan caller computing agent technology. However, nan latest announcement makes nary mention of this requirement. The institution is now moving to further heighten nan technology, aiming to support aggregate GPUs installed crossed aggregate servers successful information halfway environments.

Fujitsu said that various companies person been testing nan computing agent middleware since May 2024. Tradom, a Japanese fintech venture, has reportedly implemented nan exertion successful production, while unreality supplier Sakura is evaluating its imaginable for optimizing information halfway operations.

Fujitsu emphasizes nan worth of assets optimization successful reducing nan power depletion of GPU-based AI applications. With generative AI services continuing to predominate nan tech landscape, enterprise-grade GPUs stay among nan astir in-demand hardware components. The institution suggests that making these systems run much efficiently is important for addressing nan increasing demand.