AMD backs CPUs over data center GPUs for AI-RAN
AMD is promoting general-purpose CPUs as the foundation for virtual radio access networks, rather than the data center GPUs championed by Nvidia. Piotr Weglicki, the telco lead for AMD’s server business, questioned the suitability of power-hungry GPUs for mobile networks and argued that successive CPU generations can deliver more performance within the same power envelope.
The company’s latest approach eliminates a separate hardware accelerator for forward error correction, running the demanding network function in software instead. AMD advertises a processor with 84 cores, compared with 72 in Intel’s most advanced Granite Rapids product. Weglicki said the design lets operators allocate computing resources more freely between network functions, AI-RAN models and edge applications. Intel integrates an accelerator into its processors, an approach AMD has criticized as restricting hardware portability.
Samsung has been working with AMD on an accelerator-free platform that supports multicell capacity, offering a potential opening in a market long dominated by Intel. AMD has otherwise made little visible progress: Nokia is pursuing GPU-based RAN with Nvidia, while Ericsson remains heavily invested in custom silicon. Broader adoption would require vendors to develop software-only forward error correction and demonstrate power consumption competitive with custom chips and established virtual RAN systems.