AMD has defined a clear strategic path for its APU graphics roadmap in the coming years: RDNA 3.5 will continue to serve as the long-term solution for mainstream and entry-level APUs until nearly 2029. In contrast, RDNA 5, representing a true architectural leap forward, will be featured only in the higher-tier SoC and APU product lines. This strategy aligns consistently with AMD's technological progression observed over several generations of APUs.

Since the emergence of RDNA 3, the GPU architecture in APUs has focused on "mature cores enhanced with process and frequency scaling", and RDNA 3.5 is more of an adjustment rather than a complete redesign. It rebalances power consumption, frequency curves, and integration at the SoC level. This is evident in the Strix Point for Ryzen AI 300, the Gorgon Point for Ryzen AI 400, and the MAX series including Strix Halo and Gorgon Halo. The architecture remains unaltered, yet the compute units are scaled up in quantity and frequency, with high-end models reaching up to 40 CUs, favoring performance enhancements over new generation refreshes. The MAX series exemplifies this approach distinctly.

The underlying premise of this strategy is the controllability of RDNA 3.5 regarding power efficiency and usable area. Unlike discrete graphic cards, the GPU in an APU must share power consumption and cooling capabilities with the CPU, NPU, and memory controller. AMD has "tamed" RDNA 3.5 to maintain a stable power model suitable for both desktop and mobile APUs, hence its prolonged usage.

According to industry insiders, the Medusa Point "Ryzen AI 500" is expected to be the last mainstream APU generation to broadly adopt RDNA 3.5 iGPUs. Beyond this point, AMD will implement a differentiated product strategy within the same generation. Regular Medusa Points will continue using RDNA 3.5 (or modifications thereof), while higher-end options like Medusa Premium and Medusa Halo will transition to RDNA 5, resulting in a dual architecture within the same generation—a pioneering concept worldwide.

The decision to bypass RDNA 4 is not due to its unsuitability for APUs but rather its design priorities, which don't align with APU constraints. RDNA 4 was optimized for discrete GPUs with enhanced AI and light-tracing capabilities, alongside tech stacks like FSR Redstone, all demanding significant area and power that aren't viable for APUs. Instead of making compromises, AMD awaits RDNA 5 which approaches iGPU feature density and system synergy from an architectural standpoint.

Based on available information, RDNA 5 iGPUs will not integrate into the CPU Core Complex Die (CCD) but will instead exist as standalone GPU chip units. Kepler_L2, via AT4 GMD, AT3 GMD, hints at this segregation idea: with the GPU module physically separated from the CPU compute unit, yet integrated within the system through SoC-level interconnects. This design, closely aligned with the Strix Halo, advances the architecture to permit more CUs and complex frontend logic.

Reflecting on AMD's iGPU evolution within its APUs across generations: Vega focused on core graphics capability, RDNA 2 aimed at system-wide power efficiency and functionality, and RDNA 3 and 3.5 elevated iGPUs for light gaming and creative applications via increase in CUs and frequencies. The visibility of RDNA 5 in premium APUs effectively escalates "integrated graphics" from a supportive element to a standalone, load-bearing computation unit.

This strategic shift is also responsive to the competitive landscape. Intel has progressed its Xe architecture to Xe3, with future plans for Celestial (Xe3P) and Xe4, while expanding the Arc series into more SoC formats. For AMD, maintaining only RDNA 3.5 for all APUs hinders high-end product competitiveness in graphics performance and features. The exclusive introduction of RDNA 5 into high-end APUs highlights product tiering.

Ultimately, this strategy is not about exhausting "old architectures to the limit", but about long-term planning for cost efficiency, yield, and positioning. RDNA 3.5 will dominate the mainstream market, while RDNA 5 focuses on high-end technology demonstration and competition. Significant changes will first manifest in premium SoCs, separate from the widespread models.