Nvidia’s AI-on-5G Innovation Lab, announced this week with Google Cloud, promises to jumpstart the development of artificial intelligence (AI) workloads for 5G networks.

The lab, which will start development later this year, combines Google’s Anthos hybrid-cloud platform with Nvidia-certified hardware and software. The goal is to provide a consistent platform for developers as they build out the services and applications at the 5G edge required to make things like smart cities and factories a reality.

“We believe every industry will be transformed in the next 10 years. This is because the forces of AI and 5G connectivity are combining with the digital automation to drive the fourth industrial revolution,” said Ronnie Vasishta, SVP of Telecoms at Nvidia. “Enterprise edge, AI platforms will be implemented on general purpose cloud native solutions and will be orchestrated and managed in the same way the cloud is today.”

Anthos is built on Kubernetes, and it supports workloads running on Google Cloud Platform, its competitors' clouds including Amazon Web Services and Microsoft Azure, as well as those in on-premises data centers and at the edge. Nvidia will also provide its AI Enterprise software suite, applications, AI frameworks, and pre-trained models that take advantage of CUDA cores running in the chipmaker’s GPUs.

“In this lab, industrial companies, system integrators, and network operators will be able to develop and test their AI-on-5G enterprise applications on Google Anthos using Nvidia AI infrastructure,” Vasishta said.

The AI-on-5G hardware is built on three components: a GPU that handles layer-1 virtual radio access network (RAN) processes and AI workloads, a BlueField-2 DPU, which serves as a standardized open RAN interface and offloads the 5G user plane function, and a host CPU.

Nvidia plans to ingrate its DPU and GPU into a single accelerator card called the BlueField-3 A100, which will be available next year. And by 2024, Nvidia aims to condense all three parts into a standalone chip.

Nvidia claims its AI-on-5G hardware will enable the creation of high-performance 5G RAN and AI applications to manage emerging use cases like robotic manufacturing, autonomous vehicles, drones, and surveillance.

“We have now brought the power of AI cloud to the 5G connected enterprise,” Vasishta said. “This brings tremendous untapped monetization opportunities to operators that have already spent billions of dollars on acquiring spectrum.”

Nvidia also announced several upgrades to its HGX line of servers to address industrial AI and high performance computing (HPC) workloads.

The chipmaker’s HGX servers can now be equipped with PCIe-based A100 GPUs with up to 80 gigabytes of video memory, Nvidia’s NDR 400 Gb/s InfiBand networking cards, and Magnum IO GPUDirect Storage software. Additionally, Nvidia extended support for Arm processors, like Ampere’s Altra to the HGX family, allowing the chipmaker’s partners to build systems with the CPU best suited for the application.

Unlike the company’s DGX A100 servers, which pack as many as eight A100 GPUs into a single chassis, Nvidia’s HGX line are manufactured by its OEM partners, which include the likes of Dell Technologies, Hewlett Packard Enterprise, Lenovo, and Atos, just to name a few.

Today’s update doubles the video memory offered by each PCIe-based GPU from 40 gigabytes to 80 gigabytes of high-bandwidth memory, which boosts the overall memory bandwidth of the card by 25%, the company claims.

If this sounds familiar, Nvidia boosted the amount of video memory on its full-size A100 GPU to 80 GBs back in November. Now, customers running smaller form factor servers can benefit from the larger pools of memory.

According to Nvidia, large pools of video memory are essential in AI training workloads like recommender models. These models often have massive tables representing billions of users and billions of products.

By doubling the available video memory, Nvidia claims it can run larger workloads on a A100 PCIe card. Or, on the flip side, divvy up more memory per virtualized GPU instances. By taking advantage of Nvidia’s multi-instance GPU technology, each A100 can now be segmented into 7 GPUs, each with 10 gigabytes of video memory.

Along with more video memory, the HGX systems now feature Nvidia’s NDR 400 Gb/s InfiniBand networking for server-to-server communication in distributed workloads.

“The DPU includes acceleration engines for data security, for networking operations, for storage management, and so forth,” said Gilad Shainer, SVP of networking at Nvidia. “We are offloading all of the infrastructure management into the DPU.

These cards can be paired with Nvidia’s Quantum-2 fixed-configuration switches, which offer 64 ports, each at 400 Gb/s, or 128 ports at 200 Gb/s for a total throughput of 25.6 Tb/s.

Nvidia claims the modular switches can scale up 2048 ports at 400 Gb/s and twice that for 200 Gb/s connectivity.

“We actually provide with NDR a variety of switch systems that can help build any scale of an AI or cloud native supercomputer from a small box,” he said.

Finally, Nvidia’s Magnum IO GPU Direct Storage platform enables direct memory access between GPU memory and storage, reducing latency and enabling the system to benefit from the full bandwidth offered by the network adapters, the company claims.

“It basically allows the storage to move directly to the GPU without having to traverse the CPU,” explained Dion Harris, lead product manager of accelerated computing at Nvidia.

Nvidia’s HGX servers, which will be provided by Atos, will power the Tursa DiRAC supercomputer at the University of Edinburgh in the UK.

DiRAC stands for distributed research using advanced computing and is part of the UK’s integrated supercomputing facility for theoretical modeling and HPC-based research. DiRAC specializes in astronomy, cosmology, particle physics, and nuclear physics.

Tursa will feature 448 Nvidia A100 GPUs, with four InfiniBand HDR 200 Gb/s NICs per node. When it comes online later this year, Tursa will accelerate research into subatomic particles using data collected from experiments like the Large Hadron Collider.

“DiRAC is helping researchers unlock the mysteries of the universe,” Shainer said. “Our collaboration with DiRAC will accelerate cutting-edge scientific exploration across a diverse range of workloads.”