AMD has pulled back the curtain on its next-generation graphics architecture, revealing the technical foundations that will power its upcoming Radeon RX 9000 series. The RDNA 4 architecture represents a significant evolution in AMD's GPU design philosophy, with a renewed focus on gaming performance alongside substantial improvements to ray tracing capabilities and AI processing.
A Gaming-First Architecture
AMD's RDNA 4 has been designed from the ground up with gamers as the primary audience. Unlike previous generations that attempted to balance gaming with professional workloads, RDNA 4 concentrates specifically on high-end gaming performance. The architecture introduces heavily optimized compute units that deliver up to 2x improvement in rasterization performance compared to RDNA 2, with even more substantial gains in specialized workloads like ray tracing (2.5x) and machine learning (3.5x) per compute unit.
New Compute Engine Design
At the heart of RDNA 4 sits a redesigned Compute Engine featuring Dual SIMD32 Vector Units and Enhanced Matrix Operations. These units support multiple precision formats including 2x-16b and 4x-8b/4b dense matrix rates, along with 4:2 Structured Sparsity for doubled processing rates. A notable innovation is the dynamic register allocation system, which allows shaders to request registers from a shared pool as needed and release them when work is completed. This approach significantly improves memory latency handling and increases the overall efficiency of the shared core.
Third-Generation Ray Tracing
Perhaps the most impressive improvement comes in the ray tracing department, where AMD has implemented third-generation ray tracing units that double ray intersection rates compared to RDNA 3 at equivalent clock speeds and bandwidth. The new ray accelerators feature improved BVH (Bounding Volume Hierarchy) compression that reduces memory requirements to less than 60% of what RDNA 3 needed, thanks to a new 8-wide design. Additional enhancements include hardware instance transforms, improved RT stack management, and Oriented Bounding Boxes that can reduce traversal costs by up to 10%.
Path Tracing Capabilities
AMD is also positioning RDNA 4 for the future of rendering with path tracing support. While ray tracing casts single primary rays for reflections, shadows, and refractions, path tracing accounts for all possible light paths, creating more photorealistic scenes at a higher computational cost. To make this practical, AMD is implementing Neural Supersampling and Denoising techniques specifically optimized for path tracing workloads, similar to NVIDIA's approach with games like Cyberpunk 2077 and Alan Wake II.
FSR 4: AMD's AI-Powered Upscaling Solution
Alongside the hardware announcement, AMD revealed FSR 4, its next-generation upscaling technology. Unlike previous FSR versions that used spatial or temporal algorithms, FSR 4 leverages machine learning through the new FP8 processing capabilities in RDNA 4. Early demonstrations at CES showed significant improvements over FSR 3.1, particularly in performance mode where previous versions struggled with image quality at low render resolutions.
 |
|---|
| Comparison of upscaling methods highlighting the improvements introduced with AMD's FSR 4 technology |
FSR 4 Implementation and Performance
At launch, FSR 4 will be integrated at the driver level, automatically upgrading any game that supports FSR 3.1 to use the new AI-based upscaling algorithm. AMD claims FSR 4 in performance mode delivers a 65% performance uplift over native 4K rendering across seven tested games, with some titles like Ratchet & Clank: Rift Apart seeing performance double. The technology will be available in over 30 games at launch, including major titles like Kingdom Come: Deliverance 2, Spider-Man 2, and Call of Duty: Black Ops 6.
Hardware Specifications
The flagship RDNA 4 GPU, codenamed Navi 48, features 53.9 billion transistors on a 356.5mm² die manufactured using TSMC's 4nm process. The chip contains four shader engines, each housing eight Dual Compute Units (DCUs) for a total of 64 Compute Units and 4,096 stream processors. The architecture includes 64 Ray Accelerator engines and 128 Matrix Acceleration Engines distributed across the shader engines. Memory support includes up to 16GB of GDDR6 memory at speeds up to 20 Gbps on a 256-bit bus interface, complemented by up to 64MB of third-generation Infinity Cache.
Media and Display Enhancements
RDNA 4 also brings improvements to media encoding and decoding capabilities, with up to 25% quality improvement in AVC/H.264 encoding, 11% better HEVC encode quality, and double the AV1 throughput. The updated Radiance Display Engine supports DisplayPort 2.1a and HDMI 2.1b outputs, while enhanced FreeSync Power Optimization modes deliver lower idle power in multi-display configurations. AMD has also introduced Radeon Image Sharpening 2, which works across all APIs through a single toggle to deliver higher quality images.
The Road Ahead
While AMD has made significant strides with RDNA 4 and FSR 4, challenges remain in competing with NVIDIA's established ecosystem. FSR 4's quality improvements are promising, but the technology starts with a significant gap to close against NVIDIA's DLSS 4, which already supports over 70 games. For FSR 4 to succeed long-term, AMD will need to maintain a strong commitment to expanding game support and continuing to refine the algorithm's quality. Nevertheless, RDNA 4 represents AMD's most focused gaming architecture to date, with substantial improvements across rasterization, ray tracing, and AI processing that should make the Radeon RX 9000 series competitive in the high-end gaming market.