Graphics cards are the foundation of today’s visual and computing world. What you see on your monitor when you play games, create 3D visualisations or process AI depends on how cleverly their architecture and performance are designed. In this context, AMD RDNA architecture refers to the internal design of the chip and therefore how the compute units, caches or memory controllers are arranged and how these components work together to process data.
The AMD RDNA architecture was introduced in 2019, replacing the older GCN (Graphics Core Next) concept. It has since become the foundation of all Radeon gaming graphics cards and has gone through multiple generations. In 2025, RDNA 4 debuted, and there is already the first information about RDNA 5, which is expected to arrive in late 2026.
What is the AMD RDNA architecture and how does it work
The AMD RDNA architecture is based on compute blocks, referred to as Compute Units (CUs). Within each block are shaders, which are actually small processors designed to handle graphics tasks such as lighting, texture application and shadow generation. Their work is supervised by a front-end that distributes the individual tasks so that the graphics card is used as efficiently as possible.
A key change from the older GCN architecture was the way RDNA handles work waves (wavefronts). While the GCN used wave64, the AMD RDNA architecture added native waveform processing with 32 threads (wave32). This brought greater flexibility and more efficient use of compute units, as smaller wavefronts are better suited to diverse graphics tasks.
The cache hierarchy is also of great importance. In addition to the classic L0, L1 and L2 levels, AMD introduced Infinity Cache in later generations – a large cache integrated directly into the chip. This cache serves as a fast data stack, so the GPU doesn’t have to reach into slower GDDR-type memory as often. The result is faster response and higher performance for graphics-intensive tasks.
The entire graphics processing process takes place in a pipeline – a chain of steps where data is sequenced from geometry processing through lighting and texturing to rasterization and final image output. The AMD RDNA architecture has optimized this process to achieve higher performance with lower power consumption. It is the improved power per watt ratio that has become one of the most important pillars that differentiated AMD from the previous GCN era.
AMD RDNA generations
Since its debut in 2019, the AMD RDNA architecture has undergone rapid evolution. Each generation has brought changes to the GPU core, the way memory works, and support for new technologies. This has seen AMD gradually move from more efficient use of compute units, to hardware ray tracing, to chiplet design and advanced AI accelerators. In the following sections, we’ll explore how the different generations of RDNA architecture differed and what they brought to users.
RDNA 1 (2019)
The first generation of AMD’s RDNA architecture arrived in July 2019 with the Radeon RX 5700 and RX 5700 XT graphics. It was manufactured using TSMC’s 7nm process and delivered over 50% more performance per watt over the previous GCN architecture. A new feature was wave32 working wave processing, which meant that the compute units were able to process smaller batches of threads more efficiently (while wave64 support was retained). The shaders ran at clock speeds around 1.8 GHz, and PCIe 4.0 interface support appeared for the first time. The lack of hardware ray tracing remained a weakness, so AMD couldn’t compete with the first generation NVIDIA RTX at the time.
RDNA 2 (2020)
In November 2020, AMD launched the second generation – the AMD RDNA 2 architecture – along with the Radeon RX 6000 series. The latter brought hardware ray tracing and the first version of Infinity Cache with up to 128 MB of capacity, which compensated for the relatively narrow 256-bit bus and made data faster. RDNA 2 also became the basis for the PlayStation 5 and Xbox Series XS consoles, ensuring its rapid expansion. It brought better efficiency, higher clock speeds, and finally decent performance in ray tracing, which gamers on AMD cards lacked until then.
RDNA 3 (2022)
The third generation – AMD’s RDNA 3 architecture – debuted in December 2022 and brought a revolutionary chiplet design(splitting the GPU into multiple smaller chips instead of one large monolith). The GPU was split into a Graphics Compute Die (GCD) manufactured on a 5nm process and Memory Cache Dies (MCD) on 6nm. This move allowed AMD to reduce manufacturing costs and better scale performance across models. The flagship RX 7900 XT and RX 7900 XTX ran at clock speeds above 2.4GHz, with the ray tracing units being 50% faster than in RDNA 2. The number of front-end units was doubled and improved shaders delivered a noticeable performance boost. The downside was higher inter-chip latency, but the price/performance ratio remained very competitive.
RDNA 4 (2025)
The fourth generation – AMD’s RDNA 4 architecture – arrived in the first half of 2025 with the Radeon RX 9000 series. AMD returned to a monolithic core design, focusing on higher efficiency and strengthening specialized units.
Third-generation ray tracing accelerators provide up to twice the performance of RDNA 3, and second-generation AI accelerators bring support for both INT8 and FP8, which is important for upscaling and AI tasks. The 3rd generation Infinity Cache reaches 64 MB, works with a 256-bit bus and GDDR6 memories with up to 20 Gbps transfer rates.
The RDNA 4 architecture added PCIe 5.0 support, more advanced DisplayPort 2.1a and HDMI 2.1b display standards, and enhanced multimedia blocks with AV1 encoding including B-frames.
The flagship RX 9070 XT model with 16GB GDDR6 delivered an average 40-42% higher performance over the RX 7900 GRE when gaming at 4K resolution. In ray tracing titles, the difference was even greater. This confirmed that AMD can compete with NVIDIA even in 2025.
Multi-die MCM (middle): This is multiple chips bundled together in a single package. It represents a kind of intermediate step that allows multiple smaller parts to be combined, thus reducing costs, although the interconnections between them introduce slight latency.
Chiplet design (right): The chip is divided into specialized smaller blocks (such as compute units and I/O part) that are interconnected by high-speed interfaces. The advantage is better performance scalability and more flexible manufacturing. AMD’s RDNA 3 architecture became the first gaming graphics to take advantage of this approach.
RDNA 5 (2026-2027, expected)
The AMD RDNA 5 architecture hasn’t been officially unveiled yet, but leaks are already hinting at the first details. A slight increase in performance per core, improved AI units and more efficient ray tracing are expected. Alpha Triton-branded prototypes are mentioned – the smaller Alpha Triton 4 with 128-bit LPDDR5X memory and 24 compute units for the mid-range, and the more powerful Alpha Triton 3 with 48 compute units, 384-bit LPDDR6 memory, and expanded L2 cache for the high-end segment. According to the available information, AMD RDNA 5 architecture could arrive in 2026-2027, but the specifications are still subject to change.
AMD RDNA Architecture – Technology
AMD RDNA architecture is thus not just about the number of compute units or core clock speeds. What makes these graphics cards really interesting are the technologies that combine both hardware and software. Together, they ensure that games look realistic, run smoothly and yet the graphics card works efficiently. Each generation of RDNA has introduced new capabilities that have gradually pushed Radeon to be more competitive against both NVIDIA and Intel.
Hardware innovations
Infinity Cache is a large on-chip cache that reduces the load on the memory bus. This means the GPU doesn’t have to constantly communicate with slower external VRAM and data is available instantly. RDNA 4 uses the third generation Infinity Cache with 64 MB, which improves performance especially at high resolutions.
AI accelerators bring specialized performance for machine learning, image upscaling or other AI computations. RDNA 4 adds two per compute block, capable of working with INT8 (8-bit integer precision) and FP8 (8-bit floating precision) formats. This means several times higher performance for AI tasks and improved image quality.
Ray tracing accelerators in RDNA 4 have reached the third generation. They can calculate reflections, shadows and global illumination faster, making scenes more naturally and realistically lit.
Shaders and front-end control the processing of graphics tasks such as texturing, shading or lighting. Already RDNA 1 introduced wave32, which means processing smaller batches of threads compared to GCN, which used wave64. This change brought more flexibility and better use of compute units.
The memory subsystem and PCIe 5.0 are the basis for fast communication. RDNA 4 uses a 256-bit memory bus, GDDR6 memory with transfer rates up to 20 Gbps, and an advanced PCIe 5.0 interface that enables higher throughput between the CPU and GPU.
The multimedia units in RDNA 4 bring support for AV1 encoding with B-frame technique and enhanced image processing. This means higher quality streams and more efficient options for content creators.
Software innovations
Hardware alone is not enough, because you need software technologies that take full advantage of it. That’s why with RDNA, AMD isn’t just offering a chip, but a suite of features that improve performance, image quality and the gaming experience.
FidelityFX Super Resolution (FSR) is AMD’s flagship software solution. It works by having the GPU render the scene at a lower resolution and FSR computes it at a higher resolution. The result is a higher FPS count while maintaining image quality. The biggest advantage of FSR is its openness – it works on AMD, NVIDIA and Intel graphics.
Radeon Anti-Lag minimizes the response time between input and screen response. It optimizes communication between the CPU and GPU and reduces time, which is especially critical in fast-paced competitive games.
Radeon Boost dynamically adjusts resolution according to camera movement. In fast-paced scenes, it lowers the resolution, taking the pressure off the GPU and increasing FPS. During static moments, the image returns to full resolution.
Radeon Chill automatically manages FPS according to activity. During idle or static images, it reduces GPU power consumption and temperature, and unlocks full power during action.
Smart Access Memory (SAM) allows Ryzen processors to access the entire VRAM of the graphics card instead of a small portion of it. This direct access delivers a noticeable performance boost in games, especially at higher resolutions.
Together, these innovations form a complete ecosystem that enables AMD RDNA architecture to offer not only raw performance, but also advanced features. As a result, gamers get higher FPS, more realistic visuals and a smoother experience, while content creators benefit from multimedia enhancements.
Let’s take a look at the latest Radeon RX 9000 series graphics cards that use AMD RDNA 4 architecture in practice to give you an idea of what the design of these cards is like as well.
AMD RDNA architecture and the competitive battle
The graphics card market today is extremely competitive and the statistics speak clearly. In the second quarter of 2025, NVIDIA held approximately 94% of the market, while AMD dropped to 6%. Only a year earlier, AMD was around 12%, a significant year-on-year decline. While the overall market for dedicated GPUs grew by 27% quarter-on-quarter, AMD benefited only marginally from this growth.
This was due to the slow uptake of AMD’s RDNA 4 architecture, as the series kicked off with only a pair of RX 9070 and RX 9070 XT models. However, these suffered from limited availability and pricing that exceeded the recommended level. The mid-range cards didn’t arrive until later, so they haven’t had time to show up in the Q2 stats yet.
Looking at the competition, it is clear that NVIDIA RTX has long maintained its leadership in areas that are key to both the gaming and professional segments – ray tracing and artificial intelligence. It leverages dedicated RT Cores and Tensor Cores that enable the deployment of technologies such as DLSS. This form of upscaling builds on neural networks and can deliver higher performance without any visible loss of quality.
AMD responds with its own FidelityFX Super Resolution (FSR) solution. Its biggest advantage is its openness, it works not only on Radeon, but also on graphics from NVIDIA or Intel. Although it still lags slightly behind DLSS in output quality, the versatility and compatibility make FSR a popular choice for developers and gamers alike.
Intel has also entered the competitive fray with the Arc series. Its XeSS technology is comparable to FSR, but the driver ecosystem still suffers from shortcomings and the hardware has not yet gained enough trust from users.
Despite NVIDIA’s dominance in the high-end segment, AMD’s RDNA architecture maintains its reputation as the price/performance leader. This is why many gamers and content creators still choose Radeon graphics cards. They offer balanced performance, affordable pricing, and open technologies that don’t leave users locked into one ecosystem.
AMD RDNA Architecture – The Most Common User Issues
Every technology has its weaknesses, and the AMD RDNA architecture is no exception. Although it has brought many innovations and moved Radeon graphics cards generations ahead, users on forums, Reddit or social networks often point out certain shortcomings. Most of the problems are related to software optimization and drivers rather than pure hardware bugs.
Stuttering and ray tracing problems
For RDNA 4 cards, stuttering is mentioned several times in games built on Unreal Engine 4 when ray tracing is enabled. This is tearing/chopping of the image that spoils the smoothness of the gameplay. Reviews agree that the problem does not arise directly in AMD hardware, but rather in the fact that the engine is better optimized for NVIDIA RTX cards from the start. The temporary solution tends to be to disable ray tracing or use upscaling until the developers or AMD bring a fix in the drivers.
Higher idle power consumption
RDNA 3 users have noticed increased idle power consumption, especially when running multiple monitors or playing video. The graphics card did not always switch efficiently between power profiles. AMD has gradually addressed this issue through driver updates that better manage refresh rates.
Driver stability
Although AMD has made significant advances in recent years, there are still reports of driver instability and thus crashes, freezes or “black screen” during updates. On Reddit and forums it is therefore often recommended to use the so called “clean install”(first completely uninstall all old or corrupted driver versions and only then install the latest driver version).
Weaker ray tracing compared to NVIDIA
AMD RDNA architecture has long suffered from lower performance in ray tracing titles. Although RDNA 4 has narrowed the gap over RTX cards, NVIDIA still holds the edge in games that make heavy use of ray tracing.
Despite these shortcomings, most reviews rate the Radeon cards positively. AMD is gradually addressing the weaknesses with software updates, and users are attracted in particular by the fair price/performance ratio and the openness of the technologies that work across multiple platforms.
Conclusion
In six years, AMD’s RDNA architecture has become the foundation of modern Radeon graphics cards. Each generation has pushed the boundaries – from the more efficient design in RDNA 1, to the advent of ray tracing and Infinity Cache in RDNA 2, to chiplet design in RDNA 3, to the more powerful and AI-centric RDNA 4. With this, AMD has gradually moved closer to the technological cutting edge and confirmed that it can compete in the premium segment as well.
The future will depend mainly on the areas where NVIDIA still leads – ray tracing and artificial intelligence. The expected AMD RDNA 5 architecture is expected to bring more powerful AI accelerators, more advanced upscaling technologies and a move to new memory solutions. There is even speculation about UDNA, a unified architecture that would combine gaming RDNA and professional CDNA into one universal concept. This could strengthen AMD not only in games, but also in HPC builds and professional applications.
Although NVIDIA remains the leader in ray tracing and AI, AMD has built a strong position thanks to its fair price/performance ratio and open technologies. AMD’s RDNA architecture is thus one of the most important pillars of graphics card development today, and its future looks as promising as the road it has already traveled.