Before the rise of the current artificial intelligence craze, Nvidia's graphics processing units (GPUs) were virtually unknown outside the computer gaming industry. Their ability to process data in parallel (performing multiple computations simultaneously) later attracted the attention of artificial intelligence developers, propelling Nvidia to become the world's most valuable company. 

Today, CEO Jensen Huang seems determined to apply Nvidia's chips to various fields. This year, Nvidia pledged to collaborate with Taiwan's MediaTek to develop a personal computer chip. In addition, there are GPUs for automobiles, GPUs for robotics, and even GPUs to turn homes into mini data centers. As part of this comprehensive GPU expansion plan, Nvidia is developing a chip for 6G radio.

Nvidia confirmed to Light Reading that this move marks a significant step forward in the GPU giant's "AI-RAN" strategy. Previously, Nvidia demonstrated how its Grace Hopper superchip could replace custom chips in the Radio Access Network (RAN). Grace Hopper is a superchip that combines a regular central processing unit (CPU, the Grace part) with a more powerful GPU (Hopper part). Grace Hopper and related products will take over the computing tasks in the RAN typically performed by devices or servers (commonly referred to in the industry as Central Units (CUs) and Distributed Units (DUs)). This allows the antenna-borne Radio Units (RUs)—the other half of the RAN—located on masts or rooftops, to remain untouched.

Unlike CUs and DUs, RUs were not previously a target for Nvidia or general-purpose processor manufacturers. Long before Nvidia entered the RAN field, Intel marketed its CPUs in a similar manner as an ideal replacement for RAN computing application-specific integrated circuits (ASICs). This precursor to AI-RAN—the "virtual RAN"—theoretically allows the telecom industry to achieve a higher return on investment with chips targeting a wider range of users. However, Intel has confirmed that its latest virtual RAN product, Granite Rapids, does not include RU components and has no plans to design RU components.

Massive MIMO changed everything

Why is this necessary? For the simpler 4G and 5G radio modules, Layer 1, or Physical Layer (PHY) processing—the most IT-intensive part of RAN computation—takes place in the Data Unit (DU). Anyone dissecting a Runner (RU) with a relatively small number of antenna elements will find transceivers, data converters, and digital front-ends for converting analog signals to 0s and 1s. But they are unlikely to encounter Application-Specific Integrated Circuits (ASICs), which play a crucial role in Layer 1.

Massively Multi-Interface (MIMO) technology has changed this. This technology uses a large number of antennas and is applied in today's more advanced 5G radios. In this technology, Layer 1 functions are distributed across the DU and RU to avoid performance degradation. The RU must ensure support for beamforming, a smart technology that allows signals to be precisely targeted to specific devices rather than wasting resources covering a larger area. On the hardware side, the traditional approach is to integrate an ASIC chip into the RU for beamforming and other such "lower-level PHY" network functions.

It is precisely these application-specific integrated circuits (ASICs) located within radio units (RUs) that will be replaced by GPUs. In a detailed commentary shared via email with Light Reading, Nvidia stated that this move is inevitable as radio units (RUs) become increasingly complex. A basic radio unit (RU) contains four transmitters and receivers. 5G Advanced and 6G may increase the number of transmitters and receivers to 128, meaning that the processor performance required for low physical layer (PHY) computing will be 32 times greater than it is now. With so-called ultra-high frequency MIMO technology (deployed in higher spectrum bands in 6G), radio units (RUs) with up to 1024 transmitters and receivers have even become possible. Nvidia stated, "With the introduction of ultra-high frequency MIMO, 7GHz bands, and artificial intelligence (AI) algorithms in 6G radio units (RUs), GPUs will become key to meeting computing demands."

The absence of any RU chip could limit Nvidia's development opportunities in the 5G and 6G fields. In massive MIMO technology, the Layer 1 communication between the DU and RU is typically handled by the same chip vendor. If different vendors are used, Layer 1 software developers need to maintain two different platforms simultaneously.

About Open RAN

In Open RAN, in principle, a company's Distributed Unit (DU) can connect to another company's Receive Unit (RU) via a standardized interface. However, this arrangement is actually more complex. An unnamed RAN expert stated that this also requires software developers from both sides to publicly disclose their tightly guarded algorithms. He pointed out that their reluctance to do so partially explains why multi-vendor massive MIMO technology has seen almost no commercial deployment.

But any user using Granite Rapids to support massive MIMO will be forced to use Intel processors in the DU and other processors in the RU. Intel doesn't see this as a problem. The Open RAN 7.2x interface, defined by the O-RAN Alliance, is designed to resolve any interoperability issues, meaning that "performance no longer depends on the same chips used in the DU and RU," Intel stated. Samsung, the vendor with the most deployed virtual RAN products, confirmed to Light Reading that its RU's low-PHY processing relies on its own custom-designed ASIC.

However, according to Intel's previous strict definition, using custom chips for some computation would make the product non-compliant with the fully virtualized Radio Access Network (RAN) standard. Intel experts have previously criticized Nokia for promoting its virtual RAN product, which placed all Layer 1 functions on custom chips from Marvell Technology. This was later referred to in the industry as virtual RAN with "inline" accelerators, using only Intel CPUs to handle higher-level functions. Similarly, Samsung also had to strip away some non-virtualizable Layer 1 functions.

Nvidia stated in an email that its current proposal will revolutionize this, providing vendors with "more flexible software-defined computing platforms, rather than fixed-function chips." Given the nature of custom chips—tight coupling of hardware and software—this part of RAN computing has long been closed to the broader developer ecosystem. Nvidia claims to have opened the door to all experts familiar with its CUDA software platform, which currently boasts approximately 6 million developers. Nvidia goes a step further, designing its CUDA-based RAN computing architecture, Aerial, which is freely available to anyone.

But GPUs are huge power consumers, aren't they?

However, the industry remains skeptical about introducing GPUs into the Radio Access Network (RAN). Even those who have overcome their reservations about using GPUs in the Data Unit (DU) question their economic viability in the Radio Unit (RU). The RU is estimated to account for as much as 90% of the total power consumption of a mobile network. Any factor that even slightly increases power consumption could render the RAN unfeasible, and GPUs have long been known for their high power consumption.

But Nvidia and its allies argue that the high-power-consuming GPUs for data centers cannot compare to the GPUs the company is developing for the RAN. The company claims that in constrained environments such as "cars and robots," they already have "embedded systems capable of operating at less than 100 watts of power and 100 degrees Celsius." A source suggests that GPUs installed in the RAN RU are more likely to resemble those designed specifically for gaming.

An intriguing relationship seems to have developed between Marvell and Nvidia this year. In March, Nvidia invested $2 billion in Marvell, but it seems more interested in the latter's technology in optical communications than its Radio Access Network (RAN) technology. However, Marvell is also a RAN chip supplier for Samsung and Nokia, which also received a $1 billion investment from Nvidia and announced plans to develop GPU-compatible RAN products.

Then, in late May, Marvell CEO Matthew Murphy made the following statement during a conference call with analysts regarding the latest financial results: "Marvell will enhance its existing Octeon base station processors to enable them to work directly with Nvidia GPUs, integrating AI with wireless infrastructure into a single software-defined computing platform."

He added, "This will enable telecom operators to run 5G and 6G radio workloads as well as high-performance AI applications simultaneously on the same hardware." While details are currently unclear, this all suggests that Nvidia has seen the appeal of leveraging Marvell's RAN technology, much like its reliance on MediaTek in the PC market.

Experts maintain that any general-purpose chip designed to meet multiple needs can never match the performance of a custom chip designed for a single purpose. Few disagree. However, when these chips have no users other than telecom operators, the massive investment in developing application-specific integrated circuits (ASICs) for radio access networks (RANs) becomes less compelling. According to Omdia, a sister company of Light Reading, global operators spent only $35 billion on RAN products last year. This figure has fallen sharply from $45 billion in 2022 and is not expected to rebound.

If developing ASICs for RANs becomes uneconomical, then the performance of new general-purpose chips could surpass that of the older custom chips still in use. "The key consideration is not a single power consumption figure, but rather that as RAN functions become increasingly complex and increasingly reliant on artificial intelligence, the trade-off in economics will shift towards programmable platforms that can evolve with standards and adopt flexible deployment models, rather than fixed designs optimized for a single configuration," Nvidia stated. If Intel or other manufacturers cannot provide a general-purpose chip suitable for wireless access networks, then Nvidia may be the only option.

Source: Compiled from lightreading

View more at EASELINK