At IFA 2024 held in Berlin, Germany, Jack Huynh, Senior Vice President and General Manager of AMD's Computing and Graphics Business Group, announced that the company will unify the consumer centric RDNA and data center centric CDNA architectures into the UDNA architecture, which will lay the foundation for the company to more effectively respond to Nvidia's deeply rooted CUDA ecosystem.
When AMD abandoned its GCN microarchitecture in 2019, the company decided to divide its new graphics microarchitecture into two different designs, with RDNA designed to support gaming graphics products for the consumer market, and CDNA architecture specifically designed to meet the computing centric artificial intelligence (AI) and HPC (high-performance computing) workloads in data centers. After being unified into the UDNA architecture, developers will be able to use it more easily than they do now.
AMD stated that they had made some mistakes in RDNA before, and had to reset the optimization matrix every time they changed the memory hierarchy or subsystem. Looking ahead to the future, the company is not only considering RDNA 5, RDNA 6, RDNA 7, but also UDNA 6 and UDNA 7. Therefore, to some extent, the issue of full forward and backward compatibility will be enforced, but it needs to be planned in advance.
High end chips can establish a market, but software support will determine the winners and losers. Nvidia has taught how to use its unparalleled proprietary CUDA to build an indestructible 'moat' ecosystem.
18 years ago, Nvidia started from CUDA and laid the foundation, and one of its most fundamental advantages may be the "U" (Unified) in CUDA (Computing Unified Device Architecture). NVIDIA has only one CUDA platform suitable for all purposes, which utilizes the same underlying microarchitecture to implement AI, HPC, and gaming. The platform currently has 4 million developers.
AMD will continue to rely on the open-source ROCm software stack to counter Nvidia, but this will require joint efforts from users and the open-source community, which will bear some of the burden of optimizing the stack. AMD will take measures to simplify work and accelerate the development of this ecosystem.
What changes will UDNA undergo compared to the current RDNA and CDNA splitting? AMD did not provide a detailed introduction, indicating that there is still a lot of basic work to be done. But the obvious potential pain point that needs to be addressed is the lack of dedicated AI acceleration units in RDNA. Given that AI work currently dominates in data centers and client GPUs, adding tensor support to client GPUs seems to be a key requirement.
A unified UDNA architecture is the next logical step in competing with Nvidia CUDA, but AMD still has a long way to go, including a clear timeline for launching the UDNA architecture.