RDNA4 (Chiplet architecture with enhanced compute units)

RDNA 4 is the fourth generation of AMD graphics architecture powering Radeon RX 9070 graphics cards based on the Navi 48 die. It is an evolutionary step focused on significantly accelerating ray tracing, improving energy efficiency, and introducing machine learning technologies for lossless image upscaling.

Description

Simple Description Like a new car engine: it's redesigned inside (pistons, cooling system) to go faster and use less fuel, but externally it's the same familiar car. You just notice games run smoother, and the computer generates less heat and noise.
Medium Complexity Description This is an evolution of AMD's GPU microarchitecture. Just as Zen 4 improved upon Zen 3, RDNA4 reworks the Compute Units (CU), cache subsystem, and ray tracing blocks. The main focus is not on raw power, but on using it more efficiently (IPC uplift), enhanced Ray Tracing, and the introduction of dedicated hardware for accelerating AI workloads.
Technical Description A GPU microarchitecture aimed at abandoning the high-end segment in favor of mainstream solutions based on chiplets or a monolithic Navi 48/44 die. It assumes a new Compute Unit design (RDNA4 CU) with a reworked SIMD engine, increased Infinity Cache bandwidth (likely 3D V-Cache), a 3rd-generation hardware ray tracing engine, and a dedicated Neural Processing Unit (NPU) for matrix multiplication (WMMA), implementing WMMA-type instructions at the shader level to accelerate XeSS-like upscaling technologies.

Technical Features

Top-Level Category Hardware Component Subcategory GPU Microarchitecture
Application Domain Consumer Graphics Processing Units (GPUs) Developer/Author AMD (Advanced Micro Devices, Inc.)
Predecessor RDNA3 (Radeon RX 7000 series) Successor Not officially announced yet (likely RDNA5 or UDNA)
Status Active (Announcement pending) Abstraction Level Physical silicon implementation
Clock Speed Expected in the 3.0–3.5 GHz range (Game Clock) on a 4nm process Power Consumption (TDP) Target TDP for the mainstream segment within 150–250 W (single-die solutions)
Process Node (nm) 4 nm (presumably TSMC N4P / N4X) Backward Compatibility API-compatible with DirectX 12 Ultimate, Vulkan 1.3, OpenGL
Supported Platforms/OS Windows (WHQL), Linux (AMDGPU kernel driver) Task Type Rasterization, Hybrid Ray Tracing, Matrix Multiplication (AI Acceleration)
Error Correction Mechanism ECC support for VRAM in professional versions (Radeon Pro), absent in consumer Form Factor Discrete graphics cards (Add-In Board), possible future APU integration
Competing Technologies NVIDIA Ada Lovelace (RTX 40 series), Intel Xe2 Battlemage Reason for Predecessor Displacement Improved energy efficiency and ray tracing performance per watt

Advantages and Limitations

Advantages Significantly improved ray tracing (Ray Tracing) performance is expected compared to RDNA3, the introduction of AI accelerators to the shader core level for high-quality upscaling (FSR 4), as well as excellent energy efficiency thanks to a mature 4nm process, enabling high gaming FPS without extreme heat generation.
Limitations According to AMD's preliminary strategy, RDNA4 will lack high-end solutions (Navi 31 class), ceding the top-tier graphics card segment to competitors (NVIDIA RTX 5090). This limits absolute performance and VRAM capacity to flagship configurations not exceeding the RTX 5070/5080 level. Also, the launch delay may shorten the market window before the UDNA architecture arrives.

Application Areas

Product AMD Radeon RX 8000 series graphics cards (e.g., RX 8700 XT, RX 8600 XT based on Navi 48 and Navi 44 chips), as well as mobile GPUs for gaming laptops (Radeon RX 8000M).
Application Area Mass gaming market (1080p and 1440p), esports PCs, entry-level budget workstations for rendering and AI inference tasks, next-generation gaming consoles (potentially PlayStation 5 Pro or its successor), portable gaming devices (similar to Steam Deck, ROG Ally).