Nvidia and Softbank exploring generative AI

Companies say they are collaborating on a server platform that will host vRAN as well as generative AI apps.

There’s no doubt that generative AI, its uses and implications, has been the tech story of 2023 so far. Now Softbank and Nvidia have said they are collaborating to develop a platform that will host generative AI applications as well as optimised 5G vRAN software on the same servers.

TMN reported in March 2023 about work Softbank was doing with Nvidia on the use of Nvidia GPUs to support a Mavenir vRAN, MEC workloads and core network software from within the same server and Nividia GPU. In March, Softbank said that Distributed Units (DU) from Mavenir and MEC AI applications were working on the same architecture server.  Nvidia said this enabled much higher utilisation by dynamically allocating GPU resources on demand, so that a telco could run AI and other apps on the same GPUs and servers as it runs its vRAN.

The new platform announced this week seems to build on that multi-tenant vision, with SoftBank saying it plans to roll out at new, distributed AI data centres across Japan, uses Nvidia’s GH200 Grace Hopper Superchip. Although the press release from Softbank and Nvidia mentions generative AI, there is not much insight about how this would help SoftBank’s own operations, or the apps that it might host. Instead, most of the focus was on the benefits to SoftBank of running vRAN workloads on the platform. SoftBank did mention in a passing comment that these workloads might, themselves, benefit from AI applications.

vRAN on multi-tenant platform

SoftBank said the data centres will host generative AI and wireless applications and vRAN software on a multi-tenant common server platform, using Nvidia’s MGX reference architecture with the Arm Neoverse-based GH200 Superchips. It said it expected this architecture to improve performance, scalability and resource utilisation of application workloads.

It added that its new data centers will also be more evenly distributed across its footprint than those used in the past. This will allow them to better operate at peak capacity with low latency and at substantially lower overall energy costs, the operator said.

Nvidia claims that the Grace Hopper chip with its BlueField-3 data processing units will accelerate software-defined 5G vRAN workloads without bespoke hardware accelerators or specialised 5G CPUs. Additionally, its spectrum Ethernet switch with BlueField-3 will deliver timing protocols for 5G.

This solution, an Nvidia release said, “achieves breakthrough 5G speed on an NVIDIA-accelerated 1U MGX-based server design, with industry-high throughput of 36Gbps downlink capacity, based on publicly available data on 5G accelerators. Operators have struggled to deliver such high downlink capacity using industry-standard servers.”

MGX Reference Architecture

NVIDIA MGX is a modular reference architecture that enables manufacturers and hyperscalers to build different server variations to suit a range of applications.

The architecture also incorporates NVIDIA Aerial software for software-defined, cloud-native 5G networks, enabling operators to dynamically allocate compute resources to improve power efficiency.

Generative AI apps on the edge

Alongside using the chip to accelerate vRAN workloads, Softbank said it is “exploring” creating 5G applications for autonomous driving, AI factories, AR & VR, computer vision and digital twins.

Junichi Miyakawa, President and CEO of SoftBank, said, “Our collaboration with Nvidia will help our infrastructure achieve a significantly higher performance with the utilisation of AI, including optimisation of the RAN. We expect it can also help us reduce energy consumption and create a network of interconnected data centers that can be used to share resources and host a range of generative AI applications.”