AMD has introduced a new GPU architecture strategy by merging its RDNA and CDNA architectures under the unified framework called UDNA. The initiative is designed to streamline development and enhance competitiveness against Nvidia's CUDA platform.
Announced at IFA 2024
The announcement came during IFA 2024 in Berlin, where AMD's senior vice president, Jack Huynh, outlined the integration plan. The unification will connect the consumer-focused RDNA and the data center-oriented CDNA, facilitating application development for both consumer and data center GPUs.
Speaking to Tom's Hardware, Huynh explained that while RDNA and CDNA each offered specific optimizations, their separation complicated development efforts. The goal of UDNA is to alleviate this complexity, ensuring full compatibility moving forward and backward. This will demand thorough planning and considerable endeavor.
Developer-Centric Architecture
The UDNA architecture aims to provide a uniform platform catering to both gaming and AI/HPC workloads. Huynh mentioned the ambition to grow the developer community from hundreds of thousands to millions. The transition might be challenging, but the expected long-term benefits justify the effort.
AMD plans to preserve the performance improvements achieved with RDNA and CDNA without modifying the memory hierarchy drastically. The strategy ensures that the performance gains from earlier architectures are maintained, facilitating a smooth shift to UDNA.
Competing with Nvidia's CUDA
Nvidia's CUDA has been a major advantage, providing a unified platform for AI, HPC, and gaming. AMD aims to challenge this with UDNA, leveraging its open-source ROCm software stack. Broader adaptation and optimization of ROCm are crucial for competing effectively.
UDNA is set to tackle limitations in the current RDNA and CDNA architectures, like the absence of dedicated AI acceleration units in RDNA. Nvidia's tensor cores have been a benchmark for AI performance. AMD plans to integrate similar functionalities in UDNA to enhance competitiveness.
Jack Huynh also discussed AMD's strategy for the high-end consumer GPU market, where Nvidia leads with an 88% market share. AMD holds 12% and intends to grow its footprint by focusing on lower-cost GPUs first before targeting high-end options again.
