AMD RDNA4: a modular architecture that will change the way we think about GPUs
At the Hot Chips conference, AMD presented new details about the RDNA4 architecture, which became the basis for the Radeon RX 9000 series. The main idea is a flexible, scalable SoC design that allows you to create different GPUs from one basic template.
Key features of RDNA4:
Modularity: The GPU is divided into blocks - the bottom one is responsible for I/O, display, media, power consumption, security and system management. This block can be reused to create less powerful cards without a complete redesign.
Flexible binning: AMD can disable entire shader engines or individual WGPs (Work Group Processors), and change the memory bus width in 64-bit increments (256, 192, 128) to adapt the GPU to any VRAM.
One architecture - many SKUs: Thanks to this approach, AMD can release mid-range models even if high-end variants have been cancelled (for example, Navi 4X).
Despite rumours, AMD has not released any mobile versions of the RX 9000M. Even new laptops with Ryzen AI MAX 300 do not have RDNA4 GPUs. So far, no mobile RDNA4.
An example of RX 9000 models:
| Model | GPU | CU/SP | VRAM | Bus | TBP |
| RX 9070 XT | Navi 48 | 64/4096 | 16 GB GDDR6 | 256-bit | 304 W |
| RX 9060 XT (8 GB) | Navi 44 | 32/2048 | 8 GB GDDR6 | 128-bit | 150 W |
| RX 9060 | Navi 44 | 28/1792 | 8 GB GDDR6 | 128-bit | 132 W |
What does this mean?
It seems that AMD has appreciated the chip technology that it actively uses in its Zen family of processors. This technology allows for easy and flexible power scaling, assembling both the most budgetary Ryzen 3 and the ultra-low-end Epyc and Threadripper solutions from the same parts. The modular approach to CPUs has made the company a fortune, so it seems that the company's engineers have been tasked with scaling GPUs in the same way. If this is the case, then very soon our perception of graphics chips will change significantly.