RDNA 2 AMD’s Groundbreaking Graphics Architecture

When AMD unveiled RDNA 2 on October 28, 2020, it ended nearly a decade of the company playing catch-up in the high-end GPU market. The Radeon RX 6000 series that launched on November 18, 2020, wasn’t just competitive it challenged Nvidia’s RTX 30-series at its own game, delivered a 54% improvement in performance-per-watt over the previous generation, introduced a revolutionary cache technology that reimagined GPU memory architecture, and brought hardware ray tracing to AMD for the first time.

More than that, RDNA 2 simultaneously powered the PlayStation 5, Xbox Series X/S, and eventually the Steam Deck making it one of the most widely deployed GPU architectures in gaming history, found in devices from PC graphics cards to next-gen consoles to Tesla vehicles.

This is the complete guide to RDNA 2: what it is, how it works, every major technical innovation it introduced, the full GPU lineup it powers, how it compares to what came before and after, and why it matters even in 2026.

What Is RDNA 2?

RDNA 2 (Radeon DNA 2) is AMD’s second-generation RDNA GPU microarchitecture, released on November 18, 2020, alongside the Radeon RX 6000 series desktop graphics cards. It succeeded the original RDNA architecture (which powered the Radeon RX 5000 series) and introduced a series of significant engineering advancements that repositioned AMD as a genuine competitor to Nvidia at every performance tier.

RDNA 2 is built on TSMC’s 7nm manufacturing process the same node as the original RDNA but achieves dramatically higher performance and efficiency through architectural improvements rather than a process node shrink. It represents AMD’s first architecture to include:

• Hardware-accelerated ray tracing (Ray Accelerators)
• AMD Infinity Cache (on-die L3 cache)
• DirectX 12 Ultimate full feature support
• Smart Access Memory (SAM / Resizable BAR)
• Variable Rate Shading and Mesh Shader support

The architecture powers the PC-focused Radeon RX 6000 series, but also forms the foundation of custom SoCs in the PlayStation 5, Xbox Series X/S, and Steam Deck making RDNA 2 the architecture that defined an entire console generation.

If you’re comparing GPU performance across generations or planning a gaming PC build, our gaming PC build guides for every budget cover balanced system recommendations that factor in RDNA 2 cards. And for a direct head-to-head on specific GPU models in the generation, the GTX 1650 vs RTX cards comparison shows how AMD’s architecture stacks up against Nvidia across different tiers.

The Engineering Foundation What Changed From RDNA 1

RDNA 2 was initially described by AMD as a “refresh” of RDNA 1, but that characterization significantly undersells what changed. While the fundamental building block the Compute Unit remained structurally similar, nearly everything around it was redesigned.

Compute Unit Scaling Big Navi Finally Arrives

One of the most notable aspects of RDNA 1 was what it lacked: a large, high-end die. The flagship Radeon RX 5700 XT topped out at just 40 Compute Units in a 251mm² die with 10.3 billion transistors a moderate chip for a moderate product tier.

RDNA 2 changed this dramatically. The flagship Navi 21 die that powers the RX 6800, 6800 XT, and 6900 XT houses 80 Compute Units exactly doubling the RX 5700 XT’s count in a massive 519mm² die containing over 26 billion transistors. AMD had finally built “Big Navi,” the large enthusiast-class GPU the community had been anticipating for years.

Each RDNA 2 Compute Unit contains 64 shader cores (stream processors), organized into two SIMD32 vector units per CU. Two CUs form a Work Group Processor (WGP), sharing 32KB of L0 cache. This hierarchy from individual shader cores up through CUs, WGPs, and Shader Engines is the structural backbone of how RDNA 2 distributes and processes graphics workloads.

Clock Speed Breakthrough

RDNA 2 achieved a 30% frequency increase over RDNA 1 while using the same power a staggering efficiency improvement from pure microarchitectural optimization rather than any manufacturing process change. The PlayStation 5’s RDNA 2 GPU runs at up to 2.23 GHz; desktop RX 6000 cards boost to 2.4-2.6 GHz on their own. These clock speeds were unprecedented for AMD GPUs and represented a fundamental rethinking of how RDNA’s pipeline managed power delivery at high frequencies.

Performance-Per-Watt: 54% Uplift

The headline efficiency number for RDNA 2 is a 54% improvement in performance-per-watt over RDNA 1. AMD engineers broke this down as follows: approximately 21% came from the Infinity Cache technology (by reducing how often the GPU needed to access slower VRAM), and approximately 30% came from the clock speed improvements and architectural pipeline optimizations. The remaining improvement came from other microarchitectural refinements throughout the chip.

Infinity Cache The Architecture’s Defining Innovation

If RDNA 2 has one defining feature that set it apart from every GPU that came before it, it’s Infinity Cache.

The Problem It Solves

Increasing GPU performance traditionally required one of two approaches: add more shader cores, or increase memory bandwidth by widening the memory bus or using faster memory. Both approaches consume significant power and die area. A 512-bit memory interface which would have been needed to match the bandwidth demands of the RX 6900 XT without Infinity Cache consumes massive amounts of power and requires substantial additional silicon.

AMD product technology architect Sam Naffziger stated directly: without Infinity Cache, “We were looking at the daunting prospect of having to put a 512-bit interface and all the power, area and expense associated with that.”

How Infinity Cache Works

Infinity Cache is a 128MB on-die global L3 cache in addition to the traditional L1 and L2 caches that GPUs have always possessed positioned between the compute units and the VRAM subsystem. It acts as a high-speed buffer that stores frequently accessed data much closer to the shader cores than GDDR6 memory ever could.

The Infinity Cache has a peak internal transfer bandwidth of 1986.6 GB/s dramatically faster than the 512 GB/s bandwidth of the RX 6900 XT’s 256-bit GDDR6 interface. When data is already present in the Infinity Cache (a “hit”), the GPU accesses it at near-instant speeds with negligible latency. When the data isn’t present (a “miss”), it falls back to GDDR6 which is where the performance penalty lies.

AMD measured an average 4K gaming hit rate of 58% across top gaming titles. This means roughly 58% of the time, the GPU found the data it needed directly in Infinity Cache without touching slower GDDR6 memory. The effective bandwidth the GPU “experiences” is therefore a weighted combination of the cache’s ultra-high internal bandwidth and the GDDR6’s external bandwidth.

Practical Gaming Impact

The result of Infinity Cache is that RDNA 2 GPUs punch well above their memory bus width in gaming performance. The RX 6800 XT, with a 256-bit GDDR6 bus, can match or exceed the RTX 3080’s performance at 1080p and 1440p despite the Nvidia card having a wider 320-bit bus and using faster GDDR6X memory. The cache’s benefit diminishes somewhat at 4K, where larger frame buffers reduce hit rates and the raw bandwidth of the memory bus becomes more influential.

Infinity Cache also proved critical for ray tracing. AMD stored a “very high percentage of the BVH working set” the large data structures that ray tracing requires for calculating light ray intersections directly inside the Infinity Cache, reducing latency and dramatically improving ray tracing performance over what a raw memory bandwidth comparison would predict.

Infinity Cache by GPU Model

Hardware Ray Tracing AMD’s First Generation

RDNA 2 marked AMD’s debut in hardware-accelerated ray tracing. Nvidia had introduced dedicated RT cores with Turing (RTX 20-series) in 2018; RDNA 2 brought AMD’s answer in 2020.

Ray Accelerators One Per Compute Unit

Rather than implementing separate dedicated RT cores as a distinct fixed-function unit (as Nvidia does), AMD integrated one Ray Accelerator directly into every Compute Unit. This architectural choice means that as you move up the RX 6000 stack, cards with more CUs automatically have proportionally more ray tracing hardware.

The RX 6900 XT has 80 Ray Accelerators (one per CU), the RX 6800 XT has 72, and the RX 6800 has 60. Each Ray Accelerator handles the hardware-accelerated portion of ray tracing specifically the BVH traversal and ray-triangle intersection calculations while the standard shader code in the compute units handles ray traversal and shading.

Infinity Cache and Ray Tracing Synergy

The Infinity Cache proved particularly valuable for ray tracing. Ray tracing is extremely memory-intensive because the BVH data structures which describe the geometry of the scene for ray intersection testing can be very large. By keeping the BVH working set inside Infinity Cache rather than constantly accessing GDDR6, RDNA 2’s ray tracing performance was substantially better than its raw compute numbers would suggest.

First-Gen Limitations

Being AMD’s first hardware ray tracing implementation, RDNA 2’s ray tracing performance was generally behind Nvidia’s second-generation Ampere RT implementation in the RTX 30-series. AMD targeted 1440p as its ray tracing performance goal and that target was largely met, with the RX 6800 XT delivering playable frame rates in ray-traced games at 1440p. However, 4K ray tracing at high settings pushed beyond what first-gen Ray Accelerators could sustain.

Driver updates post-launch improved the situation significantly. By February 2023, driver updates had boosted ray tracing performance by up to 40% using DirectX Raytracing a remarkable post-launch improvement that demonstrated how much headroom remained in RDNA 2’s hardware.

DirectX 12 Ultimate Full Feature Parity With Console

RDNA 2 achieved full compliance with the DirectX 12 Ultimate specification the complete set of next-generation rendering features Microsoft defined for the new console and PC generation. This made RDNA 2 the first AMD architecture to support:

Variable Rate Shading (VRS): Allows the GPU to render different parts of a frame at different shading rates lower quality in peripheral areas the eye doesn’t focus on, higher quality at the center. This improves performance without a noticeable quality reduction in most scenarios.

Mesh Shaders: A new programmable geometry pipeline stage that replaces the older vertex shader model for certain workloads, enabling more efficient rendering of complex geometry.

DirectX Raytracing (DXR): Hardware-accelerated ray tracing through the standard Microsoft API, ensuring broad game compatibility.

Sampler Feedback: Allows games to access data about what textures were sampled during rendering, enabling more intelligent texture streaming from storage.

The combination of these features, shared between RDNA 2-based PCs and the PS5/Xbox Series X, created a unified development target that made it significantly easier for game developers to optimize for both platforms simultaneously a benefit AMD explicitly cited as a strategic advantage of being the GPU provider for both Sony and Microsoft.

Smart Access Memory (SAM) CPU-GPU Synergy

Smart Access Memory, also known as Resizable BAR (Base Address Register), is an AMD feature that gave RDNA 2 an additional performance advantage when paired with AMD Ryzen 5000 processors.

Traditionally, CPUs can only access 256MB of a GPU’s VRAM at a time a limitation of the PCIe specification’s address aperture. With Smart Access Memory, AMD bypassed this limitation by giving the CPU direct access to the GPU’s full 16GB VRAM pool simultaneously.

The practical result is reduced memory fragmentation within the VRAM pool, lower latency for CPU-GPU data transfers, and measurable performance improvements in games AMD claimed and demonstrated gains of several percentage points in supported titles. SAM was initially exclusive to the Ryzen 5000 + RX 6000 pairing, but was later supported more broadly as Resizable BAR became a standard PCIe feature adopted industry-wide.

The Full RDNA 2 GPU Lineup Radeon RX 6000 Series

Desktop GPUs

Mobile GPUs (RX 6000M Series)

Launched May 31, 2021, the RX 6000M series brought RDNA 2 to laptops with the RX 6600M, 6700M, and 6800M. These maintained the same architecture with power tuning appropriate for mobile thermal envelopes.

Console Implementations

PlayStation 5: Custom RDNA 2 SoC with 36 Compute Units running at up to 2.23 GHz. Fixed-function hardware for backward compatibility with PS4 titles.

Xbox Series X: Custom RDNA 2 SoC with 52 Compute Units at 1.825 GHz, delivering 12 teraflops of compute performance.

Xbox Series S: Custom RDNA 2 SoC with 20 Compute Units at 1.565 GHz, targeting 4 teraflops.

Steam Deck (2022): Custom “Van Gogh” APU with 8 RDNA 2 Compute Units paired with Zen 2 CPU cores, optimized for handheld power efficiency.

Additional deployments: Samsung Exynos 2200 (mobile SoC), AMD Ryzen 7000’s integrated display engine (2 CU RDNA 2 iGPU for display output), and Tesla Model S/X entertainment systems (confirmed June 2021 by both AMD CEO Lisa Su and Tesla CEO Elon Musk).

RDNA 2 Competitive Performance How It Stood Up Against Nvidia

At launch, the RDNA 2 flagship RX 6800 XT matched or beat the RTX 3080 in traditional rasterization performance at 1080p and 1440p the first time an AMD flagship had genuinely beaten an Nvidia flagship at equivalent tiers in years. The Infinity Cache’s bandwidth advantage was most pronounced at these resolutions, where its cache hit rates were highest.

At 4K, the picture was more mixed. As resolution increases, the GPU needs to process larger frame buffers that don’t fit as neatly in Infinity Cache, reducing hit rates and leveling the bandwidth advantage. Nvidia’s wider memory buses and faster GDDR6X memory gave it an edge in 4K heavy workloads.

In ray tracing, RDNA 2’s first-generation implementation generally trailed Nvidia’s second-generation Ampere RT cores, though the gap narrowed significantly through driver updates. AMD made no apologies it had publicly targeted 1440p as its ray tracing goal, and that target was met.

For competitive gaming at 1080p where RDNA 2’s Infinity Cache hit rates were highest and its clock speed advantage was fully realized RDNA 2 was exceptional. Our RTX 4060 Ti vs RTX 3070 comparison shows how subsequent Nvidia generations continued to push performance forward, which contextualizes just how much ground RDNA 2 had helped AMD reclaim.

FidelityFX and AMD’s Upscaling Answer

One area where RDNA 2 was explicitly behind Nvidia at launch was AI-based upscaling. Nvidia’s DLSS uses Tensor cores dedicated machine learning accelerators not present in RDNA 2 to reconstruct high-quality frames from lower-resolution inputs with remarkable results.

AMD’s answer was FidelityFX Super Resolution (FSR), introduced in June 2021. Unlike DLSS, FSR uses spatial upscaling algorithms that run on standard shader hardware, meaning it works on any GPU from any manufacturer. This was both a strength (broader compatibility) and a limitation (lower quality than DLSS at equivalent modes, particularly in areas of fine detail and motion).

FSR supported a wide range of RDNA 2 hardware and improved significantly through subsequent versions FSR 2.0 introduced temporal upscaling for significantly better quality, and later FSR 3.0 added frame generation.

RDNA 2’s Legacy What It Changed for AMD

RDNA 2 achieved several things that AMD had not managed in years:

Restored high-end competitiveness. The RX 6800 XT beating the RTX 3080 at 1440p was the first time AMD had a product that could genuinely make an enthusiast choose it over Nvidia’s top offering.

Defined a console generation. Powering PS5, Xbox Series X/S, and Steam Deck simultaneously meant RDNA 2 is the architecture that defined the current gaming console generation. Every developer optimizing for those consoles was optimizing for RDNA 2 features ray tracing via DXR, mesh shaders, variable rate shading.

Established Infinity Cache as a differentiator. The on-die cache approach proved so effective that AMD carried it forward into RDNA 3 (with a next-generation version). It fundamentally changed how AMD thought about the relationship between memory bandwidth and gaming performance.

Proved AMD’s efficiency trajectory. The 54% performance-per-watt improvement over RDNA 1 demonstrated that AMD’s RDNA roadmap had genuine momentum. RDNA 3 subsequently delivered over 50% additional efficiency gains over RDNA 2.

RDNA 2 vs RDNA 3 vs RDNA 4 The Architecture Evolution

Understanding RDNA 2 is incomplete without seeing where it fits in AMD’s broader architecture roadmap.

RDNA 3 was the first architecture to use chiplet design in GPUs, with the Graphics Compute Die (GCD) fabricated on TSMC 5nm and memory cache dies on 6nm. RDNA 4 (powering the RX 9000 series launched in 2025) returned to a monolithic design and introduced 3rd-generation ray tracing with 2x the throughput of RDNA 3.

AMD confirmed in 2026 that RDNA 4 is the final generation of the RDNA branding. The successor architecture will abandon the RDNA name entirely, introducing “Radiance Cores,” “Neural Arrays,” and a universal compression engine focused heavily on AI and next-generation graphics scenarios.

RDNA 2 Cards in 2026 Are They Still Relevant?

In 2026, RDNA 2 desktop cards remain functional for 1080p and 1440p gaming, though they’re increasingly showing age against modern AAA titles that stress VRAM capacities and ray tracing hardware.

The primary limitation affecting RDNA 2 midrange cards (RX 6600 XT, RX 6600) is their 8GB VRAM a capacity that’s become tight for modern AAA titles at medium-to-high settings. The 16GB variants at the top (RX 6800, 6800 XT, 6900 XT) remain capable at 1440p and hold up reasonably at 4K.

For budget builds and used market purchases, RDNA 2 cards represent strong value propositions particularly the RX 6700 XT (12GB, 40 CUs) and RX 6800 (16GB, 60 CUs), which remain solidly mid-range performers at 1440p.

The console implementations remain fully current every PS5 and Xbox Series X/S sold is still running RDNA 2, and both consoles continue to receive new game releases with no deprecation on the horizon.

For guidance on which current GPU to pair with your system, our breakdown of AMD vs Nvidia GPU architectures comparison examines how both companies’ architectures have evolved through 2025-2026.

Quick Technical Reference RDNA 2 at a Glance

FAQs

RDNA 2 stands as one of the most consequential GPU architectures AMD has ever shipped. It simultaneously won back the high-end PC GPU market, defined an entire console generation, introduced Infinity Cache as a genuinely new approach to GPU memory design, and brought hardware ray tracing to Radeon for the first time. Even in 2026, its influence is everywhere in the hundreds of millions of PS5 and Xbox Series consoles running it, in the Steam Decks still being sold and played, and in the foundation it laid for every AMD GPU architecture that followed.