AMD RDNA 4’s Modular SoC Nature & Flexible Configurability Helps Spawn Smaller & Diverse GPUs Such as Navi 44, Highlights Features To Reduce Memory & Bandwidth Needs

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

AMD RDNA 4 GPU Architecture Recapped Along With New Modular SoC Details At Hot Chips 2025

We already shared the full deep-dive into the AMD RDNA 4 architecture when it was unveiled back in February of this year. Today, AMD provides a bit more detail, and before diving into the modular SoC nature of the chip, let's talk about some new highlighted features.

One interesting point highlighted by AMD was the use of LPDDR memory for lower-end RDNA 4 GPU SoCs. They state that while LPDDR offers lower power, it still doesn't provide sufficient bandwidth, and the shoreline required for such a chip will be larger, thus increasing its package size. Hence, this is why LPDDR isn't a viable option for graphics cards.

2 of 9

On being questioned about the lower memory bandwidth versus RDNA 3, AMD stated that memory bandwidth depends on the workloads, and they have done enough tuning on the graphics architecture with RDNA 4 that helped reduce the requirement for bandwidth needs.

The new details for RDNA 4 GPU architecture, which AMD presented at Hot Chips, are about the Modular SoC architecture. AMD's RDNA 4 was designed as a flexible SoC chip, spawning a range of configurations that can be used by them in Radeon products. This section was presented by AMD's SOC Architect, Laks Pappu, for Instinct Data Center GPUs and earlier had also led the SoC architecture team for RDNA 4 and its follow-on series. So yeah, it looks like we are looking at a similar modular SoC approach on RDNA 5 / UDNA series as detailed below.

First, we have the Data Flow chart, where we get to look at the multiple Shader Engines within Navi 4X SoCs. The Shader Engine houses multiple WGPs or Work Group Processors, each with dual Compute Units.

These communicate with the memory controller and LLC using a GL2 Cache on the GPU side and the new & improved Infinity Fabric (Coherent interconnect). There are several Coherent Stations within the SoC alongside the LLC and Dual-channel memory controllers. The memory controllers are connected to the DRAM on the PCB (GDDR6). The Infinity Fabric operates at 1KB/clock (bandwidth) with a frequency range of 1.5-2.5 GHz.

Now, coming to the Modular SoC design, which has the ability to spawn smaller SoCs. AMD has drawn a red line to showcase where the modularity of the chip cuts off and can be expanded to various SKUs. The chip in the diagram below the red line is what is configured for the Navi 44 with two Shader Engines and four GDDR6 (128-bit) memory controllers, along with the rest of the IO/controllers. And it can be done both ways, scale up or scale down the SoC architecture. Maybe we will see bigger SoCs in the next-gen of RDNA GPUs.

You can simply add more SEs, L3, Infinity Fabric interconnects, and GDDR memory controllers to expand the chip to higher-end SKUs such as Navi 48, which is used in the top-tier RX 9070 XT graphics card. This modularity also enables higher security levels, allowing access to controls and different levels of privileges for component security controllers, power management, and MPI (u-controller). RAS is embedded into various components of this modular die.

AMD also dives a bit into the RDNA 4 SoC central compression and decompression algorithms. With new techniques, AMD's RDNA 4 GPUs offer 15% performance improvement in some raster workloads, deliver 25% reduction in fabric bandwidth, which leads to lower power, and obviates the need for software to be cognizant of compression algorithms (fully contained in HW).

Once again, AMD talks abflexiblelexibile configurability of its Modular SoC design for RDNA 4 GPUs. The modular architecture provides AMD with fusing options to derive multiple product SKUs based on market requirements. They have four harvest levels, which can result in various configurations:

• SE Harvest
• WGP Harvest
• Asymmetric Harvest (May Include weighted PS distro as well as weighted CS distro)
• Memory device harvest (Single device granularity / 64b granularity)

Currently, AMD has four Navi 48 SKUs and three Navi 44 SKUs available, but the modular SoC nature of RDNA 4 can enable future configurations.

Follow Wccftech on Google to get more of our news coverage in your feeds.