This article details AMD's RDNA4 GPU architecture, the foundation for the RX 9000 series, which prioritizes efficiency gains over chasing peak performance. The architecture introduces significant improvements in raytracing and machine learning while advancing rasterization, media, and display capabilities. Key architectural changes include a larger 8 MB L2 cache and the removal of the L1 data cache to better handle latency-sensitive workloads. Compute efficiency is enhanced through new scalar floating-point instructions and a "split barrier" synchronization mechanism that reduces thread stalls. The media engine delivers higher-quality, low-latency video encoding, while the display engine introduces a hardware-based sharpening filter and drastically reduces multi-monitor idle power consumption. RDNA4 leverages improved transparent compression across the SoC, allowing it to achieve high performance with a modest 256-bit memory bus. AMD's decision to use a monolithic die for this generation underscores a pragmatic design philosophy based on performance targets and cost-effectiveness.
AMD RDNA4 Architecture: A Deep Dive into Efficiency and Design
Overall Strategy: Efficiency First: RDNA4 is designed as a major efficiency leap in raytracing, machine learning, and rasterization, rather than an attempt to capture the highest-end performance segment. It powers the RX 9000 series of GPUs.
Media Engine: Higher Quality Streaming: The hardware-accelerated media engine improves video encoding quality, especially in latency-constrained modes used for streaming. Testing shows it achieves higher visual quality scores (VMAF) at lower bitrates than its predecessor, RDNA3.5, particularly with difficult content like on-screen text.
Display Engine: Power Savings and Sharpening: RDNA4 introduces two key display engine improvements:
Multi-Monitor Idle Power: It dramatically reduces power consumption in multi-monitor setups by dynamically lowering refresh rates when the screen content is static.
Radeon Image Sharpening: A new sharpening filter is implemented directly in the display hardware, allowing it to be used on any content (games, desktop) with no performance impact and better power efficiency than shader-based methods.
Compute Architecture: Smarter Thread Management:
Scalar FP Instructions: The scalar unit can now execute some floating-point instructions, which saves power and frees up the more complex vector units for other tasks.
Split Barriers: The traditional barrier instruction is split into separate "signal" and "wait" actions. This allows a thread to perform independent work after signaling it has completed a task, reducing stalls and improving overall GPU efficiency.
Memory Subsystem: Larger L2, No L1: RDNA4 features a significant architectural shift in its cache hierarchy.
8 MB L2 Cache: L2 cache capacity is increased to 8 MB (up from 6 MB in RDNA3). This larger cache is crucial for improving raytracing performance by significantly reducing slow memory accesses.
L1 Cache Removed: The mid-level L1 data cache present in prior generations has been removed, with AMD likely allocating the transistor budget to the more impactful L2 cache expansion.
Improved Compression: The architecture uses enhanced transparent compression throughout the chip (caches, display, media engine). This technology is key to reducing memory bandwidth requirements, enabling RDNA4 to achieve high performance with a relatively narrow 256-bit GDDR6 memory interface.
SoC Design: Monolithic and Resilient:
Monolithic Die: AMD chose a monolithic design for RDNA4 after evaluating performance goals and costs, concluding it was the right fit for this specific graphics engine size. This is a reminder that chiplets are a tool, not a universal solution.
RAS Features: RDNA4 incorporates Reliability, Availability, and Serviceability (RAS) features, including error correction and the ability for the driver to re-initialize parts of the SoC to recover from hardware errors without a system shutdown.
Key Takeaway: Validated Efficiency: The design choices result in impressive real-world performance. The RX 9070 offers rasterization performance similar to the previous-generation RX 7900XT while leading in raytracing, all within a lower power budget and on a smaller, more efficient die.
Error: value error Invalid operation: The response.text quick accessor requires the response to contain a valid Part, but none were returned. The candidate's finish_reason is 1.