AMD RDNA 4 – An architecture for gamers? Innovations in compute units, ray tracing and AI

After much speculation, AMD has finally released the architecture details for RDNA 4. The Radeon RX 9000 series relies on an enhanced compute structure, optimized ray tracing cores and some AI improvements. Any surprises? Not really. But at least there are a few sensible changes. If you didn’t want to or couldn’t read the long article yesterday (link at the bottom of the page) because it was perhaps too detailed, here is a short version.

RDNA 4 – progress or just evolution?

After RDNA 3 and the slightly revised 3.5 version, RDNA 4 no longer offers any high-end models, but instead some adjustments that are primarily intended to appeal to gamers.

Technical key points:

• Optimized for modern gaming workloads
• Improved raster and compute performance
• Significantly increased ray tracing performance level
• Higher memory bandwidth efficiency
• AI features for upscaling and rendering
• Multimedia improvements for streaming and encoding

According to AMD, the new GPUs should offer a doubling of rasterization, a 2.5-fold increase in ray tracing and a 3.5-fold increase in machine learning (FP16 calculations) per compute unit compared to RDNA 2. Marketing figures? Let’s see how it performs in practice.

The new compute units and ray tracing enhancements

At the heart of RDNA 4 are the revised compute units with dual SIMD32 architecture and improved matrix operations. There are:

• 2x 16-bit and 4x 8-bit/4-bit matrix rates
• 4:2 structured sparsity for double efficiency
• New 8-bit float data types
• More efficient register management for less memory latency

AMD has improved on the ray tracing side. The third generation of ray tracing units brings:

• Double cut point rates for box and triangle tests
• Improved BVH compression
• Accelerated ray processing with optimized memory management
• Hardware-assisted instance transformations

Optimized memory accesses and BVH8 structures are expected to reduce memory requirements by up to 40%. According to AMD, traversal costs have been reduced by a further 10% thanks to more efficient bounding box algorithms. A real leap or just a computing trick? Benchmarks will show.

Innovations in the cache and memory interface

The command processing has been revised, Infinity Cache has been increased to 64 MB (3rd generation), accompanied by 8 MB L2 cache and 2 MB aggregate CPU cache. AMD continues to rely on GDDR6 memory, this time with up to 20 Gbps and 16 GB capacity with a 256-bit interface. There are also new compression methods to make better use of the available bandwidth.

Path tracing: the holy grail or old wine?

Ray tracing was yesterday, today it’s all about path tracing. NVIDIA shows how it works with Cyberpunk 2077 and Alan Wake II, and AMD wants to follow suit. RDNA 4 introduces its own neural supersampling and denoising process to make path tracing more efficient. Whether this will be enough to compete with NVIDIA’s AI-based ray tracing techniques remains to be seen.

Media and display improvements

There are some adjustments in the area of streaming and encoding:

• 25% better quality for H.264 encoding
• 11% HEVC quality improvement
• B-frame support for AV1 encoding
• 30% faster encoding at 720p
• Optimization for OBS, FFMPEG and Handbrake
• 50% better energy efficiency for AV1/VP9 playback

The display output now features DP 2.1a, HDMI 2.1b and a new version of Radeon Image Sharpening 2. The new Radiance Display Engine ensures improved scaling and sharpening.

Conclusion: Revolution or evolution?

RDNA 4 brings solid improvements, but no radical change. Ray tracing performance increases, AI upscaling is integrated and memory management is improved. Whether this is enough to seriously compete with NVIDIA remains to be seen. At the latest when the first tests are available, it will become clear whether RDNA 4 is really a big step – or just a well-packaged update.

Source: AMD

AMD Radeon RX 9070 XT und RX 9070 im Detail vorgestellt, das Benchmark-Schaulaufen kommt später