Nvidia aims guns at Intel with GPU for vRAN play

Nvidia explores GPU acceleration for vRAN with Ericsson; deploys with Red Hat on containerisation.

Nvidia and Ericsson have announced they are exploring the potential to use Nvidia GPUs for vRAN processing. The companies said that they would explore if and how vRAN physical layer processing could be carried out in GPU memory, to accelerate the processing of the time-sensitive and processor-intensive vRAN workload.

The announcement came as part of a broader series of announcements from Nvidia aimed at the mobile operator and carrier space.

The chip developer announced a new edge platform, tied to partnerships with a series of virtualisation and cloud infrastructure providers to design solutions for the network edge, including AI and vRAN workloads.

But it is the work with Ericsson that attracts attention of those following vendor strategies in vRAN. That’s chiefly because Ericsson has been a staunch proponent of the value of its own ASIC-based RAN baseband processing.

The discussion on vRAN processing has been: can you do the really time sensitive and intense processing work in a general purpose x86 environment that combines FPGAs with CPU cores, or is the vRAN physical layer necessarily best suited to dedicated signal processing? There has been a related architectural debate: how and where the radio technology stack can be split, giving rise to the split Baseband Units known as the Central Unit (CU) and Distributed Unit (DU).

On the first point, the major RAN players such as Huawei, Ericsson and Noka that build their own dedicated baseband chips have said that you need that dedicated hardware. Indeed it has provided their differentiation.

Ericsson: “nothing to announce yet”

Per Navinger, Head of Product Area, Networks, Ericsson, told TMN that the company is in the earlier stages of exploring the virtualisation of the lower RAN layers.

“It’s quite an interesting point in time right now, the discussion on how we can virtualise the RAN. Of course we at Ericsson deliver on purpose-built hardware and typically purpose-built is more effective and efficient. But we continuously explore and evaluate other alternatives as to how networks can be built in the future. Everyone agrees you cannot build on pure x86 server solutions – you need something that is accelerating protocol stacks, and then we see different ways of handling this acceleration. In the Nvidia discussion we are looking at how you can use GPUs with x86 and a PCI express (PCI-e) interface to build a fully virtual RAN.

“We announced we are exploring this, and are interested to look into it, but we have not announced any products. Nvidia on their side are releasing the EGX separate from this, looking at how GPUs could be used in any type of configuration.” That refers to using the super-computer for edge applications as well as vRAN.

“The Intel alternative vRAN architecture – FlexRAN – runs part of the stack on x86 and then designs PCI-e interface to FPGAs for acceleration – that’s another way of doing it. What we are seeing is given very large bandwidths and massive mimo technology that is coming, part of the challenge is to send data across this interface. We have seen benefits of doing it in GPU and are also exploring doing it in FPGA or another processing unit,” Navinger said. 

Navinger said that nobody had yet cracked a means of virtualising the whole stack.

“We already have virtualisation of the higher layers, but it could be even more interesting if we can virtualise the whole stack. There’s no truly successful cases where this has happened* and the reason is so much processing is needed that is latency sensitive – and if you push it over those interfaces it becomes challenging. Where the GPU is programmable and very powerful – that’s where it starts becoming truly interesting.”

(* Companies such as Altiostar and Mavenir might dispute this characterisation, especially for 4G.)

Nvidia’s Aerial takes to the stage

Nvidia’s proposal is to put the vRAN baseband physical layer processing in GPU memory. Notably, its recent acquisition of Mellanox provides it with the accelerated hardware path between the network interface and the GPUs, reducing the potential latency involved in accessing the parallel processors in the GPUs.

Nvidia has also introduced an SDK – Aerial – that will allow its EGX edge computing platform to be used as part of a vRAN deployment.

The company said, “With Nvidia Aerial, the same computing infrastructure required for 5G networking can be used to provide AI services such as smart cities, smart factories, AR/VR and cloud gaming.”

Aerial provides two SDKs — CUDA Virtual Network Function (cuVNF) and CUDA Baseband (cuBB) — to give developers the means to deploy 5G RAN network software to COTS servers with its GPUs.

The company said: “The cuVNF SDK provides input/output and packet processing, sending 5G packets directly to GPU memory from GPUDirect-capable network interface cards.

The cuBB SDK provides a GPU-accelerated 5G signal processing pipeline, including cuPHY for L1 5G Phy, delivering unprecedented throughput and efficiency by keeping all physical layer processing within the GPU’s high-performance memory.”

The Nvidia EGX stack includes a driver, Kubernetes plug-in, Container runtime plug-in and GPU monitoring software.”

Nvidia has said it will be using RedHat Openshift for its Kubernetes deployment.

Red Hat addressing operator demands

Azhar Sayeed, Chief Architect, Service Providers, Red Hat, told TMN, “FPGA and GPUs are widely proposed in Intel servers for handling the DU functionality (L1 and L2 processing) to reduce the CPU burden. For example, FPGA/GPU coupled with a 40 core server can easily handle 24 cell sectors.”

It is enabling that CU-DU split that we mentioned before that is driving many of the lower layer virtualisation efforts.

“Every operator we have spoken to and every design we have reviewed required us to add FPGA/GPU at the DU layer to process L1 and L2 in the RAN stack to make it economical for deployment of the RAN,” Sayeed added.

Sayeed said that Red Hat is working with partners to containerise the L1, L2, L3 and RIC [RAN Intelligent Controller – an O-RAN specified component -TMN] for vRAN so that they can run on OpenShift, its Kubernetes environment.

“We are working with Intel, NVIDIA and others to enable the hardware functionality, including acceleration, and expose that via Kubernetes so that RAN applications can make use of them.”