AMD’s RDNA 4 Strategy: Two Dies, Many GPUs, Zero Waste

▼ Summary
– AMD’s RDNA 4 architecture features asymmetric harvesting, allowing flexible scaling of GPU performance by selectively disabling components like shader engines and memory controllers.
– This approach enables AMD to create multiple product variants from the same base design, including higher-end, mid-tier, and specialized GPUs, improving manufacturing efficiency and yield.
– Memory harvesting allows incremental reduction of bus width in 64-bit steps, while compute-to-pixel resource ratios can be adjusted to tailor products for gaming, multimedia, or compute tasks.
– Asymmetric harvesting helps AMD reduce costs, shorten time-to-market, and respond to market needs by maximizing usable dies from each wafer and maintaining consistent features across product lines.
– AMD has built seven desktop and server products using just two RDNA 4 processors (Navi 48 and Navi 44), though it opted not to pursue the high-end desktop GPU market with this architecture.
Modern graphics processing units are engineered with remarkable flexibility, allowing performance to be scaled both during the design phase and after manufacturing. AMD’s RDNA 4 architecture, also referred to as Navi 4, takes this concept further through a strategy called asymmetric harvesting. This approach enables the company to produce a wide range of GPUs from just two base designs, maximizing efficiency and minimizing waste.
At the heart of this strategy lies the ability to selectively disable components within the GPU. Each chip contains numerous similar elements, compute units, cache structures, memory controllers, and specialized hardware. If a defect occurs during production, the affected section can be shut down, and the chip can still function reliably. AMD has designed RDNA 4 to be especially adaptable, allowing not only defect recovery but also intentional performance tiering.
By reusing photomasks from larger designs, AMD significantly reduces manufacturing expenses. For instance, the smaller Navi 44 die, which powers the Radeon RX 9060-series, is derived from the larger Navi 48 design. Key reductions include fewer shader engines, less Infinity cache, and a narrower memory interface, while components like the command processor, display engines, and media blocks remain fully intact. This method not only cuts costs but also accelerates time-to-market by reducing the need for separate validation and production cycles.
A central element of this approach is the modularity of the shader engine. Each shader engine houses multiple work group processors, compute units, and fixed-function hardware. With RDNA 4, AMD can disable entire shader engines when defects are present or when targeting a lower performance bracket. Additionally, finer control is possible by turning off individual work group processors, enabling unusual configurations like the 56-compute-unit Radeon RX 9070.
Memory subsystem flexibility is another critical advantage. The memory controllers in RDNA 4 are linked via Infinity Fabric and can be disabled in 64-bit increments. This allows AMD to adjust the effective bus width depending on the product tier. For example, the flagship Radeon RX 9070 XT uses a full 256-bit interface, while the RX 9070 GRE steps down to 192-bit. Even mid-range models like the RX 9060 series can be configured with 128-bit buses, supporting different memory capacities based on market demand.
Beyond sheer performance scaling, asymmetric harvesting allows AMD to tailor products for specific workloads. By adjusting the ratio of compute resources to pixel processing power, the same silicon can be optimized for gaming, multimedia, or compute tasks without architectural changes. This granular control extends to the system-on-chip level, where shared components like L2 cache, compression hardware, and Infinity Fabric links maintain efficiency across all configurations.
Reliability and security are baked into the design, ensuring that even partially disabled dies perform flawlessly. Robust error handling means AMD can confidently market these chips as fully functional products, regardless of how many components are active. This approach not only improves yields but also allows the company to address multiple market segments with fewer unique designs.
From a business perspective, asymmetric harvesting has enabled AMD to create seven distinct desktop and inference server products using only two GPU dies: Navi 48 and Navi 44. This efficiency could easily extend to mobile variants if AMD chooses to compete more aggressively in the notebook GPU space. Although the company has opted not to challenge the very high-end desktop market with RDNA 4, the architecture’s flexibility provides a strong foundation for future generations.
The lessons learned from RDNA 4’s asymmetric harvesting will undoubtedly influence upcoming architectures like UDNA. As AMD continues to refine its manufacturing and design strategies, consumers can expect even more efficient and versatile graphics solutions in the years ahead.
(Source: tomshardware)