Long before the current situation of Nvidia’s GPUs holding AI hostage, the company tried to put its chips in mobile devices but failed. The Tegra mobile chip proved CEO Jensen Huang can fail, but he seemingly never gives up.
Nvidia’s now trying another way to get into mobile devices: by putting its GPUs in equipment that form the foundation of 5G networks.
The company announced it was working with a Japanese telecom provider to put GPUs across its 5G network. Nvidia’s GPUs are being installed in waypoints as data moves through networks, providing AI on the edge to deliver smarter data to devices.
The edge GPUs spread performance across the network. It reduces the load on mobile devices and data centers to run AI tasks.
Nvidia’s equipment fits into OpenRAN designs, which are being used by service and equipment providers to design hardware for telecom networks. OpenRAN ensures a smooth data flow and equipment interoperability between servers, base stations, and edge devices.
All major server, cloud, and chip providers participate in OpenRAN designs with the expansion of 5G networks. Dell and HPE provide server designs, and Google Cloud certifies equipment to connect to their cloud services.
Nvidia is plugging its Aerial platform into networks, including the A100X GPU and Bluefield networking equipment. The equipment runs AI applications in virtualized environments. The DoCoMo implementation uses real-time operating systems, in this case, software from Wind River.
The real-time nature means the dynamic allocation of virtualized AI environments depending on a network’s load. The AI applications can be developed using the CUDA platform, though Nvidia’s GPUs support a range of open-source AI frameworks.
Nvidia’s closed and proprietary GPU designs and software infrastructure are an unwelcome addition to the open-source ethos of OpenRAN. In this case, Nvidia equipment on the edge may be inoperable on hardware at other waypoints and base stations. That depends on the input and output mechanisms and whether Nvidia’s equipment is self-contained at each base station. Many server makers have introduced proprietary designs to sit within OpenRAN networks, but are usually based on open-source software.