SynaXG shows carrier-grade AI-RAN on shared NVIDIA GPUs

SynaXG demonstrated concurrent 5G FR1, 5G FR2 and AI workloads on a shared NVIDIA GPU platform at Mobile World Congress Barcelona 2026, aiming to show how operators could deploy radio access networks.

The Singapore-based company said the set-up delivered carrier-grade 5G throughput and latency while running AI workloads simultaneously. The demonstration used NVIDIA's AI Aerial platform and a single NVIDIA GH200 system, with GPU resources shared across RAN and AI tasks.

Single platform

Radio access networks have often relied on dedicated hardware for baseband processing, with separate infrastructure for AI and analytics. SynaXG is positioning its approach as software-defined RAN on accelerated computing infrastructure, with the option to pool compute resources across functions.

In the MWC demonstration, SynaXG said it ran 20 x 100 MHz 5G NR cells and delivered more than 36 Gbps of aggregated throughput with sub-10 millisecond latency. It also said the system supported up to 1,200 connected user equipment per cell.

The company said performance was comparable with leading commercial 5G deployments, while keeping radio network key performance indicators stable during concurrent AI processing.

mmWave milestone

SynaXG also presented what it described as the first carrier-grade FR2 virtualised RAN running alongside FR1 and AI workloads on shared GPU infrastructure. FR2 is commonly associated with millimetre-wave spectrum, which operators and enterprises have explored for dense urban coverage and private networks where very high capacity is needed.

It reported end-to-end latency as low as 5 milliseconds in the FR2 set-up, saying the results met performance requirements for dense urban and enterprise deployments.

Virtualising FR2 workloads is generally seen as more demanding than lower-band configurations, given the throughput targets and tight timing requirements in the RAN. If repeatable in operational environments, SynaXG's claims would point to a broader range of RAN functions moving to shared accelerated infrastructure.

Orchestration focus

A key part of the demonstration was SynaXG's real-time orchestration software, which it said dynamically allocates and switches Multi-Instance GPU partitions based on live RAN and AI performance indicators. The approach slices the GPU and then rebalances those partitions across workloads as conditions change.

According to SynaXG, GPU resources were rebalanced across FR1 coverage functions, FR2 capacity functions and AI workloads without service interruption. It presented this as a way to maintain deterministic RAN performance while improving utilisation of compute resources that might otherwise sit idle during parts of the daily traffic cycle.

That pooling concept matters for operators experimenting with Open RAN and cloud RAN architectures, where general-purpose compute and disaggregated software offer flexibility but can introduce performance and cost trade-offs. GPU-based acceleration is one option for addressing that gap, particularly as telecoms groups add AI inference and training requirements at the network edge.

Operational claims

SynaXG said the AI-RAN system operated continuously, 24 hours a day, 7 days a week, under sustained load, which it positioned as evidence of stability and operational resilience for carrier environments.

It also said the platform maintained predictable performance while adapting to changing traffic conditions and AI workload demand. The claim addresses operator concerns that AI processing could contend with time-sensitive radio workloads with strict latency and reliability requirements.

Alongside the MWC demonstration, SynaXG said it is extending its AI-RAN platform to additional NVIDIA systems, including NVIDIA DGX Spark, describing this as portability across NVIDIA CUDA-based platforms.

The company is also working on what it calls AI-for-RAN functions covering optimisation and automation, as well as spectral and energy efficiency. It listed initiatives spanning integrated sensing and communications, distributed MIMO and inter-cell interference coordination.

In addition, SynaXG said it has been working with partners on AI agents that analyse real-time network conditions and service demand and then take actions that affect performance and energy use.

Xin Huang, Chief Executive of SynaXG, said the results show AI-RAN can move beyond lab concepts and into deployment planning.

"With the recent industry-leading breakthroughs, SynaXG has demonstrated that AI-RAN can deliver carrier-grade FR1 and FR2 performance with continuous 24x7 operation on shared NVIDIA AI infrastructure," said Xin Huang, CEO, SynaXG.

NVIDIA framed the demonstration as support for a shift toward software-defined wireless infrastructure.

"Software-defined architecture is key to the next generation of wireless networks. SynaXG's benchmark 5G performance on NVIDIA AI-RAN platform proves that operators and enterprises can achieve the flexibility and agility of cloud-native computing while maintaining carrier-grade throughput and performance per watt, required for commercial 5G services", said Soma Velayutham, VP of AI and Telecoms, NVIDIA.

SynaXG said it will show the AI-RAN set-up as part of AI-RAN Alliance Working Group 2 demonstrations and at its own booth during Mobile World Congress in Barcelona.