New Fujitsu technology distributes GPU capacity between applications in real time

Fujitsu is introducing a new technology that can distribute computing time for both CPU and GPU in real time based on the needs of applications. In theory, this means that all available capacity is optimally used.

Fujitsu is proposing a new technology with which it wants to optimally distribute the computing power of the CPU and GPU among applications. According to Fujitsu itself, its new tool is the first that can distribute computing time in real time based on the needs of applications. The Japanese company hopes that the solution will ensure that available hardware is used more efficiently. In this way, the technology offers perspective for the global shortage of GPUs.

Optimal distribution

CPUs and GPUs can serve different applications simultaneously, but that does not mean that the capabilities of both components are optimally used by default. Fujitsu gives an example where three applications compete for the computing time of two GPUs in a server.

In such a scenario, it is possible to assign GPU-one to app one, and GPU-two to app two. When app three is launched, the software will temporarily use one of the GPUs to accelerate the application workloads. The results of that measurement are compared with previous measurements in apps one and two. It may then turn out that app three benefits more from GPU acceleration than app one, reducing the overall compute time on the server when GPU two is allocated to app three. When app two is then ready on GPU two, app one can use that freed up graphics card.

Real-time priority

In HPC systems, the software can also switch between different programs in real time, so that, for example, code that needs to be executed in real time is given priority. According to Fujitsu, this is useful for applications within digital twins, generative AI and more. In both cases, the setup is similar: software demand for GPU acceleration exceeds supply, so Fujitsu’s technology plays referee so that computing power is allocated in the most efficient way possible.

Fujitsu initially wants to implement the new technology in its own solutions. The Japanese are first looking at the Fujitsu Kozuchi AI platform, but the company’s HPC service would also benefit from it.