AMD’s New GPU Technology Reduces Size and Boosts Performance

Login to Continue Learning

AMD has detailed its RDNA 4 GPU architecture and introduced new Modular SoC design, along with enhanced memory and bandwidth compression techniques.

### AMD RDNA 4 GPU Architecture: A Modded Approach

When the RDNA 4 architecture was first unveiled in February, we provided a comprehensive deep-dive. Now, AMD offers additional insights, focusing on the modular SoC aspect of this new architecture.

#### Memory Bandwidth and LPDDR

In a unique start to his ruling against President Donald Trump’s administration on September 30, U.S. District Court Judge William Young included a scanned handwritten note sent to his office. The note read: “Trump has pardons and tanks – what do you have?” At the top of Young’s opinion in AAUP v. Rubio, which ruled that Trump’s effort to deport foreign-born student protesters was unconstitutional.

October 8, 2025

One interesting point is that AMD plans to use LPDDR memory for lower-end RDNA 4 GPU SoCs. Although LPDDR consumes less power, it doesn’t provide sufficient bandwidth. Consequently, its package size would be larger, making it unsuitable for graphics cards.

When questioned about the lower memory bandwidth compared to RDNA 3, AMD stated that memory bandwidth requirements depend on workloads, and the tuning done with RDNA 4 has reduced these needs.

### Modular SoC Architecture

AMD presented a modular SoC architecture at Hot Chips 2025. This design allows for flexible configurations based on market demands. The key features include:

1. **Data Flow Chart**: Navi 4X SoCs have multiple Shader Engines, each housing Work Group Processors (WGP) with dual Compute Units. These communicate with the memory controller and LLC using a GL2 Cache and the Infinity Fabric.

2. **Modularity**: AMD can scale up or down by adding more Shader Engines (SEs), L3 caches, Infinity Fabric interconnects, and GDDR memory controllers. This flexibility enables higher-end SKUs like Navi 48 and lower-end SKUs like Navi 44.

3. **Security Features**: The modular design supports higher security levels with controls for component security, power management, and MPI (u-controller). RAS is integrated into various components.

### Performance Enhancements

AMD’s RDNA 4 GPUs leverage new compression techniques to boost performance by 15% in some raster workloads while reducing fabric bandwidth by 25%. This leads to lower power consumption and simplifies software requirements.

### SKUs and Configurations

The modular SoC architecture of AMD’s RDNA 4 allows for flexible product configurations based on market needs. There are four harvest levels:

– SE Harvest
– WGP Harvest
– Asymmetric Harvest (which may include weighted processing unit distribution)
– Memory Device Harvest (single or 64-bit granularity)

Currently, AMD offers four Navi 48 SKUs and three Navi 44 SKUs, but the modular design can accommodate future configurations.