In the future, Nvidia wants to increasingly rely on ARM processors in the 5G and Edge areas in order to provide AI computing power in the vicinity of the RAN (Radio Access Network) in a more efficient manner . The aerial platform offered by Nvidia for network operators will be expanded accordingly to include BlueField accelerators.
Nvidia calls the solution previously offered to network operators, Aerial A100, which is also offered as a developer kit. An A100 accelerator is combined with a classic x86 CPU to provide a virtualized radio access network (vRAN) that is controlled purely by software instead of using specially programmed FPGAs and ASICs. Areas such as beamforming, modulation and demodulation, parity check and others are handled by it.
On the other hand, Nvidia provides AI computing power at the edge, i.e. in the immediate vicinity of the wireless 5G interface, in order to supply clients such as intelligent cameras or robots. This can be of relevance in public networks, such as those offered by network operators in Germany, but above all in private 5G networks, such as those operated by companies in manufacturing or retail.
BlueField-3 meets A100
For MWC 2021, Nvidia is expanding its aerial division and will increasingly use ARM in the future -Processors put. Even before the takeover of ARM is completed, the group is gaining increasing relevance in solutions from Nvidia.
With the availability of BlueField-3 in the first half of 2022, Nvidia also wants to offer a plug-in card called BlueField-3 A100 in order to offer “AI-on-5G” as a more compact and efficient solution. Compared to the previous server implementation, three times the performance per watt should be offered. Nvidia's BlueField-3 DPU (Date Processing Unit) uses 16 ARM Cortex-A78 cores and integrates the ConnectX-7 network adapter for 400 Gbit/s for the first time. The RAM is set to DDR5 and the interface for integration into the data center already uses the upcoming PCI Express 5.0.
Roadmap for AI-on-5G (Image: Nvidia) 5G is to be offered on a chip
Nvidia also gives a small outlook on the year 2024, the “AI-on-5G” with the two separate components ARM-CPU and Ampere-GPU on one card is supposed to unite to form an independent chip. For the GTC, Nvidia had announced BlueField-4 based on 64 billion transistors with a connection of up to 800 Gbit/s and 1,000 TOPS computing power. The next but one generation “AI-on-5G” could use this solution as a basis and combine it with a powerful new generation of GPUs. “Ampere Next”, “Ampere Next Next” and our own Grace CPU are already in the pipeline.