Your browser is not supported. Please update it.

9 August 2021

Nvidia’s work on VMware’s Project Monterey will filter into 5G vRAN

By Wireless Watch Staff

Work that Nvidia and VMware are doing on acceleration for distributed computing could prove important in the bid to design off-the-shelf processors that are capable of supporting the demands of a distributed 5G macro vRAN.

Along with other applications that require very high levels of processing power and very low latency, such as some AI workloads, or advanced security, the Layer 1 and 2 functions in a 5G vRAN put an enormous strain on processors. That has led to most RAN functions in early virtualized RANs remaining on the radio unit, or on a dedicated appliance.

In the emerging architectures for vRAN and Open RAN, the distributed unit (DU) is designed to support Layer 2 and some Layer 1 functions, including very demanding ones like Massive MIMO beamforming, on cloud infrastructure and merchant processors. This involves a large number of accelerators that can offload some of the processes from the CPU, but that adds to cost and power consumption. Companies such as Qualcomm, Intel, Marvell and Nvidia are working on various solutions, based on different types of chips including graphical processing units (GPUs) and FPGAs, that can support future vRAN DUs without compromises in performance or power efficiency, compared to dedicated solutions.

Nvidia has been working with Ericsson on a prototype DU solution based on its signature GPU technology. This can achieve the performance – it is used to power some of the world’s biggest supercomputers – but it is challenging to do so with the compact form factor and low power threshold required for a DU.

Now the work Nvidia – as well as Intel – is doing with VMware may fill some missing pieces of the puzzle. Project Monterey is VMware’s development of a management plane for distributed compute acceleration, and it moved into early access last week after nearly three years in the making. The project provides a consistent platform for distributed compute across multiple hardware accelerators using data processing units (DPUs) supplied by merchant chip providers such as Nvidia and Intel.

Motti Beck, senior director of enterprise market development at Nvidia Mellanox, wrote in a recent blog that the growth of complex workloads like AI and real time data streaming analysis were “reducing data center efficiency and can prevent IT from meeting its service level agreements”. DPUs aim to address that challenge by offloading network, storage or security processing from the CPU, freeing up compute resources.

The hyperscalers have been deploying DPUs for years, and have built up the considerable software engineering expertise required to optimize them for the infrastructure. These skills will be needed as telcos start to plan distributed clouds to support disaggregated 5G RANs, though some will rely on hyperscaler partners for the platforms rather than trying to build their own from scratch. Either way, they will be keeping an eye on projects such as Monterey, which appear data center-focused, but will become increasingly relevant as RANs move to the cloud.

Project Monterey works as a management plane for orchestrating compute resources using DPUs such as Nvidia’s BlueField-2. It allows users to create virtual machines using resources that are distributed across the data center rather than being limited to a single server.

Paul Turner, VP of product management for VMware’s Cloud Platform Business Unit, told SDxCentral; “By putting ESXi managed through vCenter, we can start building a distributed virtual machine. To do that, we need to have a control software that actually can manage aggregating and pooling of resources.” This is a superior approach to thee “myopic” one of using smartNICs alone for network monitoring and performance offload, he claimed.

Nvidia’s third generation BlueField DPU claims to be able to offload I/O tasks equivalent to 300 CPU cores, and to run the entire VMware ESXi virtualization stack.