Fujitsu has begun providing middleware that allows companies to fully utilize their scarce GPUs for AI. The software should provide up to 2.25 times more efficient computing power for these applications and therefore lower GPU usage costs.
Late last year, Fujitsu already hinted at the “AI computing broker,” behind which lies a GPU allocator technology for AI applications. With the recent announcement, after extensive testing earlier this year, this technology is now available to potential end users.
Optimal GPU performance
Fujistu’s AI technology distinguishes between programs that do require a GPU to work optimally and those that require only a CPU. In addition, the middleware ensures that computing power is allocated in real time to prioritize AI processes with high execution efficiency, even if the GPUs in question are already running programs.
It is also possible to combine the relevant GPU optimization software with specific AI optimization technology so that the software can identify the desired AI processes and allocate computing power for each individual GPU.
Additional memory capacity
Furthermore, via memory management hidden in the middleware, the technology appears to ensure that GPUs can simultaneously run more AI processing than the GPU’s memory can handle. Fujitsu indicates that the tech works up to 150 GB of AI processing or five times the actual memory capacity of a GPU.
Future developments
The Japanese tech giant announced that it will continue to develop its new AI technology in the near future. The goal here is to ensure that the technology will soon work across multiple GPUs in multiple servers. Currently, the technology only works for individual GPUs in single servers.
In the (near) future, Fujitsu also wants to use this technology to address challenges surrounding the running of GPUs and possible power failures. In this way, Fujitsu wants to start ensuring that AI will actually, like a faithful assistant, bring more productivity and creativity.
Fujistu’s AI computing broker technology is generally available immediately. Two major Japanese end users, Tradom and Sakura Internet, will soon start using the technology for their own applications, such as in their own data centers.
Also read: GPU shortage drives Fujitsu to make best use of existing hardware