AMD to Launch Unified UDNA GPU Architecture

At IFA 2024 in Berlin, Jack Huynh, Senior Vice President and General Manager of AMD's Computing and Graphics Business Group, announced the company's plan to unify its consumer-focused RDNA and data center-focused CDNA architectures into a single UDNA architecture. This move is seen as a strategic step towards effectively challenging NVIDIA's well-established CUDA ecosystem.

In 2019, AMD shifted away from its GCN microarchitecture and introduced two distinct designs: RDNA for gaming graphics and CDNA for data center workloads, particularly AI and high-performance computing (HPC). By unifying these architectures into UDNA, AMD aims to simplify the development process, making it easier for developers to work with a single architecture.

AMD acknowledged past challenges with RDNA, particularly the need to reset optimization efforts every time changes were made to the memory hierarchy or subsystems. Looking ahead, the company is not only focusing on future iterations of RDNA but also on UDNA, aiming for forward and backward compatibility, which requires careful planning.

While advanced chips can help capture market share, software support is critical in determining long-term success. NVIDIA has demonstrated this with its proprietary CUDA platform, which has become a robust ecosystem over the past 18 years. The unified "U" in CUDA has allowed NVIDIA to use a single platform for AI, HPC, and gaming, currently supported by over four million developers.

AMD continues to rely on its open-source ROCm software stack to compete with NVIDIA, but this approach requires collaboration from both users and the open-source community. AMD plans to streamline the development process and accelerate the growth of this ecosystem.

The exact changes that the UDNA architecture will bring compared to the current RDNA and CDNA split remain unclear, as AMD has not provided specific details. One potential challenge lies in the lack of dedicated AI acceleration units in RDNA, an area that is becoming increasingly important in both data center and client GPUs. Adding tensor core support to client GPUs could be a crucial step in addressing this need.

The unified UDNA architecture represents a logical next step in AMD's competition with NVIDIA's CUDA, but the company has a long journey ahead, including finalizing a clear timeline for UDNA's rollout.