RDNA5 explained: AMD’s next-gen Radeon GPUs target AI and high-end gaming

AMD is preparing one of the most sweeping updates to its Radeon lineup: the RDNA5 architecture, which some insiders already refer to as UDNA. According to analyst Kepler, RDNA5 represents a substantial step beyond RDNA4, which focused on mid- and high-end solutions and did not include integrated or mobile versions. The new generation targets both gamers and AI, positioning it as AMD’s most ambitious effort in years.

One of the flagship options in the series, the AT0 GPU, is expected to carry 96 compute units arranged as eight arrays of 12. This layout could deliver a strong jump in performance and push AMD closer to NVIDIA’s territory, where AD102 and GB202 still lead on raw power and memory capacity. Lower-tier chips scale accordingly: AT2 is said to offer up to 40 units with a 192-bit memory bus, while AT3 will feature 24 compute units and support for LPDDR5X.

Memory is a focal point: AT0 is tipped to use a 512-bit bus, promising massive bandwidth. Each shader block will integrate a render backend, a graphics engine, and L2 cache. Notably, AT3 and AT4 are set to employ more memory controllers, tied to LPDDR5X support and, potentially, greater flexibility in mobile devices.

Specifications are not final, and AMD has yet to produce test silicon, so these details should be treated as provisional. Even so, the direction suggests RDNA5 could mark a turning point—narrowing the gap at the top end while opening new avenues for AI. Taken together, it reads as a bold bid to regain momentum where it matters most.