Will GPU and AI Advancements Prove Most Effective in Advancing Open RAN Technology?

Generative AI has exploded onto the consumer scene in 2023, with enterprises now prioritizing the development of AI models tailored to their use case. These models can span various AI disciplines, including generative AI, Machine Learning (ML), and Deep Learning (DL). AI and ML have been used to optimize Radio Access Network (RAN) architectures through automation and predictive AI/ML models for data processing. Graphics Processing Unit (GPU) solutions have their parallel processing capability as an advantage against Central Processing Unit (CPU) solutions, as they are effective at accelerating the performance of AI/ML and, in turn, Virtualized RAN (vRAN) solutions.

In March 2023, Open RAN vendor Mavenir, in collaboration with SoftBank and NVIDIA, was able to achieve End-to-End (E2E) communication from user equipment to image processing Multi-access Edge Computing (MEC) applications with vRAN components that use GPUs on actual machines. With the advancements in GPU technology by NVIDIA, such as the Grace Hopper Superchip, potentially removing the need for custom hardware accelerators for vRAN software, Commercial-Off-the-Shelf (COTS) servers can better host both RAN and AI applications, such as generative AI. NVIDIA and SoftBank also announced, in late May 2023, that the Grace Hopper Superchip will be used to power SoftBank’s new, distributed data centers across Japan, and NVIDIA believes that the superchip will accelerate software-defined 5G vRAN. The Grace Hopper Superchip combines CPU, GPU, and Data Processing Unit (DPU) technology in one chip, allowing for performance gains for 5G vRAN with improved energy efficiency. The DPU manages the networking functionality, while the GPU currently handles the Layer 1 (L1) acceleration, and the CPU handles the Layer 2 (L2) and above of a vRAN network.

This is just the latest among a few announcements in this area. Fujitsu announced its new 5G vRAN solution in February 2023, aiming to enable telecommunication carriers to build flexible open networks. Red Hat also announced that it is working with NVIDIA to allow RAN deployments on industry standard servers and remove the need for custom hardware. Artificial Intelligence (AI) will play an important role in assisting the development of software for GPU accelerators, unlocking more of vRAN’s potential:

• Reduced Cost of Architecture: GPUs are not affordable to incorporate into vRAN architectures if they are used solely for vRAN purposes. Generative AI is being used by developers as an assistive tool in their workflows. Improved software for GPU accelerators will ensure multi-purpose capabilities and allow for GPUs to be used on multiple vendor solutions. Although this will require integration of GPUs to other solutions for vendors, leveraging improved partnerships with GPU providers, such as NVIDIA, across their portfolios will be beneficial in the long term.
• Improved Performance: Improved design of GPUs will lead to various performance benefits. Power consumption and efficiency gains will not only help resource management, but will also allow vendors to use fewer GPU accelerators to run their architecture. This will be important given the necessity to reduce Total Cost of Ownership (TCO) for vRAN architectures.
• Improved Resource Management: Improved DL models will enable extracting patterns from network traffic data more efficiently. As DL models eliminate the need for feature engineering and can identify patterns on their own, implementing deeper models, which are more computationally efficient, will provide more robust analysis of network traffic without the need for feature engineering, which is required in traditional ML models. Furthermore, DL advancements can improve GPU utilization, ensuring hardware is being maximized, which will assist in keeping the amount of hardware necessary for vRAN down.

The advancements in GPU technology can be leveraged by System Integrators (SIs) and software vendors. SIs can use GPU adoption to increase the cost effectiveness and potential performance improvements for their vRAN solutions. Software vendors can use GPUs to further improve the performance of their existing software products, such as Virtualized Network Functions (VNFs), AI and analytics software, and others, potentially leading to increased demand for their solutions. Use of GPU technology in RAN processing can lead to more flexible, scalable, and energy-efficient telco networks, enabling telcos to better respond to market changes and market demands. While this is a growing field, there is a lot of work to be done to optimize the stack to greatly outperform other conventional systems, which may be based on custom silicon. This is because, for example, the Grace Hopper Superchip is not a purpose-built accelerator, but rather a software-defined accelerated platform. NVIDIA aims to provide full-stack acceleration of RAN networks beyond L1 with this superchip solution over the coming years.

Currently, Original Equipment Manufacturers (OEMs), such as Ericsson, Huawei, and Nokia, have a dominant position in the market. However, there are still opportunities for other market players, such as hardware vendors like Fujitsu, software vendors, and SIs, to improve their market position. An effective strategy can be shifting the focus to enhancing the performance of existing AI models. Below are ways to further the adoption of AI technology and improve existing and future RAN solutions:

• AI-Driven Chip Design: Adoption of generative AI across the tool flow can be used to verify and test semiconductors, which can help chip vendors optimize performance and lower power consumption, driving their time to market.
• Leveraging Optimization Tools to Support Performance of Existing Hardware: AI can be used to monitor network performance to identify bottlenecks or inefficiencies and adjust the network parameters to improve performance, leading to improved coverage optimization and load balancing through using improved ML models.
• Predictive Maintenance: Generative AI and DL can be useful for proactive identification of network failures or malfunctions through pattern identification in historical data, coupled with maintenance requirement predictions, improving sustainability metrics across networks, and increasing uptime.
• Network Security: Open RAN solutions run on open-source software, which provides many benefits for implementation, but comes with known security risks from bad actors. Existing safeguards and security measures can be further enhanced by using ML and DL models for detection of anomalies in radio traffic.

The focus for hardware vendors should be on analyzing and deciding upon adopting a GPU-based approach for vRAN solutions. Given the novelty of the Grace Hopper Superchip and the plans for optimization of the solution for vRAN purposes, vendors must, of course, be cautious before changing course from custom silicon solutions. Adopting GPU for solutions can lead to better integration of AI technologies, such as generative AI and DL, which can potentially lead to improvements at all layers of the stack for vRAN solutions, further driving the time to market. However, it is vital for ready-to-deploy solutions and services to be provided to attract more demand in this technology domain from telcos. NTT DOCOMO has been conducting tests for 5G vRAN in June 2023 with Fujitsu, solidifying the interest telcos have in the GPU-based approach for Open RAN networks. Mass adoption of GPU technology could be the next step for vendors, but it will take time to ensure GPU technology is optimized for the vRAN use case.