Nokia, Ericsson, and Samsung Diverge on GPU Integration for 6G RAN

The evolution of radio access network (RAN) infrastructure is encountering a significant architectural divide as vendors prepare for the 6G era. While Nokia has committed to a future centered on Nvidia graphics processing units (GPUs), competitors Ericsson and Samsung are maintaining a more cautious stance, emphasizing the continued relevance of custom silicon and central processing units (CPUs).

Nokia is betting that the computational intensity of advanced machine learning algorithms, such as those involving reproducing kernel Hilbert space (RKHS), necessitates the parallel processing power of GPUs. Pallavi Mahajan, the chief technology officer at Nokia, noted that these mathematical operations were previously impractical for standard infrastructure due to their heavy reliance on vector computation.

By integrating Nvidia hardware, the company aims to move away from custom silicon entirely. It intends to treat future network updates as software-defined iterations rather than hardware overhauls.

This strategic pivot follows a $1 billion investment from Nvidia and a commitment to utilize the CUDA platform for future 6G product development. Nokia claims this transition could double spectral efficiency by 2028, a projection that has sparked debate among industry peers regarding the feasibility of such gains.

The company intends to phase out its reliance on custom silicon from partners like Marvell Technology in favor of a unified GPU-based architecture. This shift represents a fundamental departure from the traditional reliance on application-specific integrated circuits that have dominated the RAN market for decades.

Ericsson has publicly challenged the necessity of this GPU-centric model, arguing that purpose-built silicon remains more efficient for specific network workloads. A spokesperson for the Swedish vendor stated that while GPUs are one potential tool, they are not a universal requirement for AI-driven RAN deployments.

The company continues to prioritize energy efficiency and performance metrics that it believes are better served by specialized hardware designs. Technical constraints regarding latency and power consumption drive the opposition from Ericsson’s engineering team.

Michael Begley, head of RAN compute at Ericsson, highlighted the rigid transmission time interval (TTI) of 500 microseconds as a critical bottleneck for any compute architecture. Because radio conditions shift rapidly, AI models must be lightweight enough to execute within this narrow window, rendering the massive parallel muscle of a GPU potentially redundant for many standard tasks.

Samsung occupies a middle ground, viewing CPUs as the primary foundation for RAN while acknowledging the utility of accelerators for specific intensive workloads. The South Korean vendor compares the versatility of a CPU to a sports utility vehicle, capable of handling general-purpose tasks efficiently, while characterizing GPUs as specialized sports cars.

Samsung maintains that a hybrid approach, rather than a GPU-only mandate, represents the most pragmatic path forward for 6G network design. The fundamental tension lies in the trade-off between throughput and the strict timing requirements of cellular networks.

Samsung noted that maximizing GPU performance often requires feeding the hardware significant amounts of data, which risks overstepping the TTI and increasing system latency. Engineers must balance the raw computational capacity of these chips against the physical realities of signal transmission, where power-to-performance ratios remain the primary metric for operator adoption.

Industry observers are now monitoring how these diverging hardware strategies will influence future 3GPP standards and operator procurement cycles. The success of Nokia’s GPU integration will likely depend on its ability to demonstrate tangible spectral efficiency gains without incurring prohibitive energy costs.

Conversely, if Ericsson and Samsung can continue to extract performance from custom silicon, the industry may see a prolonged period of architectural fragmentation in 6G deployments. The outcome will ultimately dictate whether the future of telecommunications hardware converges on general-purpose AI accelerators or remains rooted in highly optimized, domain-specific silicon.