Today, NVIDIA announced its 5G Data Centers With Gen AI with an anchor customer, Japanese mobile operator SoftBank. NVIDIA says its rollouts start in late 2023, which is right around the corner in telecom terms; we are not sure exactly when SoftBank plans to roll it out. NVIDIA claims its new GH200 Grace Hopper Superchip will allow Radio Access Network (RAN) calculations on par with current ASIC systems from significant suppliers; however, the Superchip can be used for many other workloads at an operator. As a result of allowing non-telecom and telecom workloads, NVIDIA Superchip customers will have a shared AI computing infrastructure that can be used to process text-to-video sessions, customer chat/avatar sessions, workforce allocation optimizations, and other in-house workloads. Additionally, NVIDIA GH200 Superchip customers could also offer this GPU/AI infrastructure to customers to make Machine Learning calculations, a potential source of new revenue for an operator.
We see NVIDIA’s leverage on the telecom industry because its GH200 Superchip can be used for many workloads. Until now, the most efficient way to perform baseband calculations for RAN has been to use purpose-built ASICs – they are good at it, and that is what leading RAN equipment companies like Ericsson, Nokia, and others do. However, the AI trend is taking off now, and all types of new AI workloads are becoming apparent, all with potentially very significant positive impacts on businesses across many sectors, including telecom. We listed a few of the workloads above, but it doesn’t take a lot of imagination to see how an operator could improve its customers’ experience by leveraging AI. Once an operator builds a GH200-based Superchip computing infrastructure, the company can use this shared infrastructure to support many workloads, and that’s the magic of today’s announcement.
Generally, operator infrastructures are built for “peak” and operate at “average,” which means that to support busy times – holidays, rush hour, and the like – their systems have to be built not to fail under those conditions; that means lots of equipment. But, across the seasons and days, these systems don’t run but at a small fraction of peak traffic conditions, which generally means they operate at something like 1/4th of peak. So, the assets go wasted most of the time. By sharing computing infrastructure with other workloads, operators won’t need dedicated computing systems for mobile RAN (which most people call 5G).
NVIDIA still needs to convince companies like Ericsson and Nokia to partner. But, the announcement with Softbank is robust evidence that they should.