"Future of 5G RAN" by Nbidia DOCOMO and Fujitsu's "5G RAN Future" | Business Network.jp

"The economic impact of AI is expected to be over $10 trillion by 2035, but this cannot be achieved without the use of 5G technologies such as low-latency communications and network slicing," NVIDIA said in 2021. Online conference "NVIDIA AI DAYS" held on June 16th and 17th. Makoto Noda, his manager at Developer Relations, who is an evangelist on AI, 5G/Beyond5G for telecom carriers at the company, held a session titled "His 5G/6G & Next-Generation Communication Infrastructure Accelerated by GPU/DPU." I emphasized this. NVIDIA, one of the companies that has contributed most to the evolution of AI in recent years, sees that AI will not be able to maximize its value without integration with 5G. So now NVIDIA has started to put a lot of effort into "AI-on-5G". It is a new communication infrastructure that integrates AI, MEC (Multi-Access Edge Computing), and 5G. At NVIDIA AI DAYS, Mr. Noda explained NVIDIA's solutions for telecommunications carriers, followed by a panel discussion "Now and the Future of Open RAN" with key persons from NTT Docomo and Fujitsu. There was a lively discussion on infrastructure. This article focuses on the evolution of 5G RAN from these two sessions. A new accelerator card that embodies "AI-on-5G" As introduced at the beginning, NVIDIA believes that by combining 5G's ultra-high-speed communication and low-latency communication, AI applications will have a greater economic effect in various industries. I believe that it will bring about As a product that embodies this, NVIDIA plans to launch a new accelerator card "NVIDIA Aerial A100" from the second half of 2021. According to Mr. Noda, the GPU, which specializes in AI processing, and the DPU, which is capable of high-speed packet processing, are "integrated into a single card to speed up AI execution and 5G processing at the edge."

"Aerial A100" scheduled for release in the second half of 2021 (click to enlarge)

This will change the way RAN and MEC facilities are built. Until now, it was necessary to combine multiple dedicated servers to configure the RAN and MEC functions, but "it can be easily realized with a general-purpose server and a single card," he said. While minimizing the installation space, power consumption, and operation effort, he emphasized the advantages of "higher processing performance and expandability than ever before." The number of “AI-on-5G partners” that develop 5G RAN, 5G core, security functions, etc. using NVIDIA products such as Aerial A100 is also increasing. As shown in the image below, Ericsson, Fujitsu, Mavenir, etc. are developing solutions for 5G RAN and 5G core areas, while Red Hat, VMware, Google, Palo Alto Networks, etc. are developing solutions for 5G virtualization/cloudization and security areas.

Partners to jointly develop solutions that embody AI-on-5G

Advanced RAN with “NVIDIA + Arm” Mr. Noda also introduced that he is working with partners on various initiatives to advance 5G RAN. One of them is the joint development of the “RAN + MEC Platform” with Arm, which announced its acquisition in September 2020. NVIDIA and Arm are currently running a PoC to use NVIDIA GPUs and Arm CPUs together to realize Cloud RAN. Cloud RAN operates cloud-native RAN functions on general-purpose compute platforms in place of traditional dedicated hardware. The functions of the CU (Centralized Unit: aggregation base station) and DU (Distributed Unit: decentralized station) that make up the RAN are deployed in the edge data center, and the RU (Radio Unit: antenna section) is accommodated there for efficient operation. The aim is to operate RAN” (Mr. Noda). In order to run this Cloud RAN on a general-purpose server, the challenge is how to perform high-speed arithmetic processing. "There is a lot of layer 1/2 processing, and in addition to increasing the CPU performance of general-purpose servers, a hybrid approach that incorporates dedicated processors to perform layer 1 and beamforming processing is required."

Image of platform using Arm CPU and NVIDIA GPU together

In response to this issue, NVIDIA provides a mechanism that offloads the enormous parallel processing that occurs in layer 1 processing to the GPU, using the Arm CPU as the main axis. In addition, we are working with Arm and other 5G ecosystem partners to realize a "RAN + MEC Platform" that runs MEC applications such as AI video analysis on the same general-purpose hardware.