en.Wedoany.com Reported - Nvidia is developing a GPU chip for 6G radio to replace traditional custom silicon, marking its latest move under the AI-RAN strategy. The graphics processing unit manufacturer has confirmed to Light Reading that its goal is to enter the core component of the radio access network—the radio unit (RU).
Before the current AI boom, Nvidia's GPUs were virtually unknown outside the computer gaming industry. Their parallel processing capabilities later attracted AI developers and transformed Nvidia into the world's most valuable company. CEO Jensen Huang is committed to embedding chips into more areas. The company's PC chip, developed in collaboration with Taiwan's MediaTek, is scheduled for launch this year. Additionally, there are GPUs for automotive, robotics, and turning homes into mini data centers. As part of this strategy, Nvidia is developing a dedicated chip for 6G radio.
This move marks significant progress in the GPU giant's AI-RAN strategy. Previously, Nvidia has demonstrated how the Grace Hopper superchip can replace custom silicon in the radio access network. Grace Hopper and related products will take over RAN computing in devices or servers, known in the industry as the central unit (CU) and distributed unit (DU). The radio unit, located on antenna poles or rooftops, represents the other half of the RAN equation.
Unlike CU and DU, the RU has not previously been a target for Nvidia or general-purpose processor manufacturers. Before Nvidia entered the RAN space, Intel promoted its CPUs as an attractive alternative to application-specific integrated circuits (ASICs) for RAN computing. This virtual RAN predecessor theoretically allowed the telecom industry to use chips sold to a broader audience, thereby enjoying better returns on investment. However, Intel has confirmed that its latest virtual RAN product, Granite Rapids, does not include any RU components, and there are no plans to design one.
In massive MIMO scenarios, the situation changes significantly. This technology, involving abundant antennas, is common in today's more advanced 5G radios. In this scenario, Layer 1 functions are divided between the DU and RU, with the RU required to support beamforming. The traditional approach includes an ASIC in the RU for low-PHY network functions such as beamforming. Nvidia plans to replace these ASICs in RUs with GPUs.
Nvidia stated in an email that this move is necessary as RUs become increasingly complex. A basic radio includes four transmitters and receivers. 5G Advanced and 6G could increase this number to 128, requiring 32 times the processor capability. In ultra-massive MIMO deployed in higher 6G spectrum bands, RUs equipped with up to 1024 transmitters and receivers are now even conceivable. Nvidia claims that as ultra-massive MIMO, 7GHz, and AI algorithms are introduced into 6G RUs, GPUs will become key to meeting computing demands.
The lack of an RU chip could limit Nvidia's opportunities in 5G and 6G. In massive MIMO, the same silicon vendor tends to bridge Layer 1 between the DU and RU. Using different vendors requires Layer 1 software developers to work with two independent platforms. Samsung, the vendor deploying the most virtual RAN products, relies on its own ASIC for low-PHY processing in its RUs, as confirmed to Light Reading by the company. In open RAN, devices from different companies can theoretically be connected via standardized interfaces. The 7.2x interface defined by the O-RAN Alliance aims to address interoperability issues, with Intel stating that performance does not depend on using the same silicon in the DU and RU. However, according to an anonymous RAN expert, in practice, this requires both parties' developers to disclose tightly guarded algorithms, which they are reluctant to do, partly because almost no multi-vendor massive MIMO has been commercially deployed.
Nvidia's proposed solution aims to change this by offering vendors a more flexible, software-defined computing platform. The company claims it has already opened doors to any expert familiar with its software platform CUDA, which has approximately 6 million developers. Nvidia has also designed a CUDA-based RAN computing architecture, branded as Aerial, available for anyone to use freely.
The industry remains skeptical about introducing GPUs into RAN, especially regarding power consumption. According to one estimate, RUs account for up to 90% of mobile network energy consumption. Nvidia and its allies argue that power-hungry GPUs designed for data centers cannot be compared to GPUs being developed for RAN. The company already has embedded systems capable of operating at power levels below 100W and temperatures up to 100 degrees Celsius. According to one source, the GPU entering the RU may be more similar to those designed for gaming.
The evolving relationship between Marvell and Nvidia is also noteworthy. In March, Nvidia invested $2 billion in Marvell, while showing strong interest in the latter's optical expertise. Marvell is a RAN silicon supplier for Samsung and Nokia, with Nokia also receiving $1 billion from Nvidia and announcing plans to develop GPU-compatible RAN products. In late May, Marvell CEO Matthew Murphy stated in a conference call that the company would enhance its existing Octeon base station processors to work directly with Nvidia GPUs, integrating AI with wireless infrastructure on a single software-defined computing platform. This suggests Nvidia has seen the appeal of leveraging Marvell's RAN expertise, similar to how it leveraged MediaTek in the PC space.
Experts believe that general-purpose chips may lag behind custom silicon in performance. However, developing ASICs for RAN requires massive investment, and these chips have no other audience besides telecom companies. According to Omdia data, global operator spending on RAN products last year was only $35 billion, down from $45 billion in 2022, with no signs of recovery. Nvidia states that as RU functions become increasingly complex and AI-native, the economic trade-off will shift toward programmable platforms that can evolve with standard advancements and flexible deployment models, rather than fixed designs optimized for a single configuration. With no other company, including Intel, offering general-purpose silicon for RUs, Nvidia may be the only option.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
