AMD RDNA 4 Deep Dive: Exploring The Tech That Powers Radeon RX 9070
RDNA 4 Will Power The Forthcoming AMD Radeon RX 9070 Series And It Looks Impressive
| AMD RDNA 4 And Radeon RX 9070 Series: Radeon RX 9070 - $549, 9070 XT - $599 MSRP A look under the hood of RDNA 4 tells us AMD's Radeon RX 9070 and 9070 XT target the midsection of the PC gaming market, where the volume is, and appear poised to make a big splash at very competitive price points.
|
|||
|
|
||
After many months of rumors and wild speculation, AMD just officially unveiled its RDNA 4 GPU architecture, giving us a glimpse into the graphics engine that will power the company's forthcoming Radeon RX 9070 series graphics cards. At a high level, RDNA 4 is AMD’s latest consumer GPU architecture, which was designed to enhance efficiency over the previous generation, while also optimizing performance for today’s more-taxing gaming and AI workloads. RDNA 4 features next-gen Ray Tracing engines, dedicated hardware for AI and ML workloads, better bandwidth utilization, and multimedia improvements for both gaming and content creation.
With this latest architecture, AMD is claiming a significant generational performance uplift per Compute Unit for better rasterization, ray tracing, and machine learning performance, while also supporting forward-looking features and technologies. There’s a lot to get to, and not a lot of time to do it, so let’s get started...
Enter The AMD Radeon RX 9070 XT
By now, most of you reading this will have heard of the Radeon RX 9070 and 9070 XT and may have digested some of the many rumors that have leaked over the last few months, but now it’s time for the real deal.The Radeon RX 9070 XT and 9070 are both built around the same Navi 48 GPU – it’s just scaled back somewhat on the non-XT cards to hit a lower price point. Said GPU has a PCI Gen 5 native, 16 lane interface and is manufactured on TSMC’s 4nm node. The chip is comprised of approximately 53.9B transistors, with a die size of about 356mm2. That’s significantly smaller than the aggregate die sizes of the Radeon RX 7900 XT/XTX (~550mm2), which uses multiple chiplets – one larger chiplet for the compute die, with multiple, smaller memory controller chiplets all around. AMD's Radeon RX 9070 and 9070 XT, in contrast, use a single, monolithic RDNA 4 die.
Tunneling a little deeper, the GPU itself is arranged into 4 shader engines, each with 8 workgroup processors, which are equipped with two compute units each. This is a similar structure to previous gen, RDNA 3 GPUs, but there are many enhancements in the design, which improve performance and efficiency and add new features.
Within each CU there are updated 3rd generation Ray Accelerators that offer 2x the peak throughput of RDNA 3. RDNA 4’s Ray Accelerators offer double the ray intersection rates, with improved BVH compression, accelerated ray traversal and shading, and add support for a new feature called Oriented Bounding Boxes – more on that last one later. 3rd Generation Matrix Accelerators are also present, which offer improved performance, along with support for 8-bit float data types, with structured sparsity support as well.
The GPU features 2MB of aggregate CU cache, along with 8MB of L2 cache, and 64MB of 3rd generation Infinity Cache situated out at the edges, between the memory controllers and shader engines. The L2 cache surrounds an updated, centrally located Command Processor, with enhanced packet acceleration. The updated Command Processor also supports some new instructions and features improved branch prediction, to better predict data fetches and keep caches better utilized.
The GPU connects to 16GB of GDDR6 memory operating at an effective data rate of 20Gbps, over a 256-bit interface (four, 4x16 memory controllers). The memory controllers also feature enhanced memory compression technology, to make better use of available bandwidth.
The Radeon RX 9070 series' actual specifications are outlined above. For this launch, although they're pictured in the slide above, AMD won't be releasing cards of its own, but rather leaving all of the designs up to its board partners. The Radeon RX 9070 XT is the full implementation of the Navi 48 GPU, with 64 total CUs, 64 Ray Accelerators, and 128 AI (Matrix) Accelerators. It offers up to 1,557 peak TOPs (Int4 with Sparsity) with a boost clock of 2.97GHz and a total board power of 304 watts. The Radeon RX 9070 scales the CUs, and Ray and AI accelerators back to 56, 56, and 112, respectively, with a 2.52GHz boost clock and only a 220 watt board power. The Radeon RX 9070 will offer 1,165 peak TOPs.
If you were expecting a GeForce RTX 5090 killer, the Radeon RX 9070 series is not it. Rather, AMD is targeting more affordable price points with these cards and performance levels in the neighborhood of the GeForce RTX 4070 Ti or RTX 5070 Ti, depending on the workload.
Dual media engines and an updated Radiance Display Engine are present on the GPU as well. The media engines support H.264, HEVC, and AV1 accelerated encode and decode, and offer better output quality and increased performance (up to 30% at 720p), at lower power. There’s a roughly 25% increase in H.264 low latency encode quality, and 11% improvement with HEVC. Basically, RDNA 4’s new media engine can more quickly produce cleaner, less blocky output, from the same input data, thanks to upgrades to its motion estimation algo and multi-frame reference capabilities. AMD is also claiming a 50% performance uplift for AV1 and VP9 decode, with reduced context switching overhead and memory write access, both of which ultimately save power.
AMD's Radiance Display Engine has also been updated to enhance idle power when dual FreeSync displays are connected, and it supports the latest flip model in Windows, with hardware Flip Queue support. This allows frame pacing to be managed by the GPU, rather than the CPU for video playback.
The Radiance Display Engine also incorporates Radeon Image Sharpening 2, which is implemented prior to the output being sent to a display. RIS2 enhances the sharpness of images sent to the monitor. Because RIS 2 is now a part of the display engine, it works across all APIs and can be enabled or disabled with a single driver toggle.
AMD RDNA 4: New Features And Capabilities
AMD's RDNA 4 GPU architecture enables a number of new features and technologies, which should ultimately increase performance and efficiency and make the Radeon RX 9070 series better suited for the latest games and AI workloads.
One feature coming with RDNA 4 that should increase performance and ensure optimal GPU utilization, is dynamic registers. With RNDA 3, the shaders allocated registers for the worst case scenario, but that led to significant variance in live registers. With RDNA 4, however, shaders can dynamically allocate registers from the available pool when needed, and they are released back into the pool when work is complete. Ultimately this allows the architecture to better utilize available registers and potentially have more waves in-flight at any given time.
We mentioned another new feature, Oriented Bounding Boxes, a bit earlier in this piece. Currently, Bounding Volume Hierarchies (BVH) are commonly used acceleration structures for ray tracing on the GPU, because of their relatively low memory footprint and flexibility in adapting to temporal changes in scene geometry. With previous generations, the bounding boxes used to build the BVH were aligned to the vertical and horizontal axis in a scene, which could result in much larger boxes than technically necessary to bound the necessary geometry. With OOB in RDNA 4, however, the bounding box can be rotated and better aligned to any off-axis geometry. This results in a reduction of traversal steps and more efficient use of GPU and memory resources, which should improve performance.
RDNA 4 also supports Out Of Order memory operations. With RDNA 3, memory operations were performed in the order in which they were made, which could result in additional latency should one wave require requesting new data, while another wave is waiting on its next step. RDNA 4 allows for requests from different shaders to be performed out of order, which allows for more efficient execution and reduced latency overall.
As mentioned, RDNA 4 is a significant step forward for GPU Ray Tracing as well. RDNA 4 doubles throughput and features dedicated hardware for Instance transform and stack management acceleration. Although RDNA 4 also supports an 8-wide BVH, new primitive node compression technology actually reduces the memory requirements.
How all of these new advancements affect ray tracing performance is outlined in the slide above. The increased throughput and BVH8 support account for the vast majority of the increase, but OBB, OOO Memory, the new BVH compression, and instance transform all contribute as well.
AMD FSR 4 And HYPR-RX Updates Built For Radeon RX 9070
Arriving alongside the Radeon RX 9070 series is the latest iteration of FidelityFX Super Resolution – FSR4. Enhancements to HYPR-RX are coming as well. For those unaware, FSR 4 features ML-powered upscaling, to improve image quality and reduce common issues associated with current upscalers, like ghosting, incorrect anti-aliasing, particle artifacts, moiré effects, and more. AMD claims near native – and sometimes better – image quality with FSR 4 upscaling enabled, but obviously at much higher perceived performance, because the games are actually rendered at a lower input resolution.
Unlike previous versions of FSR, which didn’t leverage any sort of machine learning, FSR 4 benefits from an AMD-built custom game ML models and a new, RDNA 4-accelerated AI upscaling algorithm. FSR 4 was specifically developed for RDNA 4 and works alongside FSR Frame Generation and Radeon Anti-Lag latency reduction. FSR Frame Generation will inject a single generated frame at this time (though more are possible in the future). We should also mention that FSR 4 uses FSR 3.1’s upgradable API, so for any games that are already FSR 3.1 enabled, FSR 4 is essentially a drop-in update.
Which brings up to HYPR-RX. If you recall, HYPR-RX isn't a single feature, but rather refers to a toggle in AMD’s driver control panel that serves to activate multiple features at once, including Radeon Super Resolution (with or without Fluid Motion Frames 2), Radeon Boost, and Radeon Anti-Lag(+). When enabled, HYPR-RX toggles on Radeon Super Resolution, Radeon Boost, and Radeon Anti-Lag(+), but note that Radeon Boost and Anti-Lag(+) don’t work in every game. In games that aren't supported by those two technologies, however, users will still benefit from the performance improvements that come by way of Radeon Super Resolution. HYPR-RX is basically a one-click performance enhancing solution for gamers that may not be savvy enough, or simply don’t want to, manage the individual technologies on their own.
AMD’s in-driver frame generation technology, Fluid Motion Frames, is also getting updated for this launch. AMD Fluid Motion Frame 2.1 will offer increased image quality for generated frames, with reduced ghosting and better temporal tracking, especially in fast-paced action. FMF 2.1 can also better handle overlays and resolve and render fine details. FMF 2.1, however, is not an RDNA 4 exclusive. It will be available with Radeon RX 6000 series GPUs, and newer, and it’ll work on AMD Ryzen AI 300 series APUs as well.
A Wealth Of New AMD Software Adrenalin Edition Updates For RDNA 4
There are a multitude of updates coming to AMD’s Adrenalin Edition software too, many of which will be AI-infused.
The previously mentioned Radeon Image Sharpening 2 update is coming specifically for the Radeon RX 9000 series, which offers stronger and more responsive sharpening with all APIs, and even on the Windows desktop, because it’s now part of the display engine. A new toggle will allow RIS2 to be applied across the entire desktop, whereas the previous version was limited to just individual applications.
AMD is also introducing a new AI Apps Manager. You can think of the AMD AI Apps Manager as a one-stop-shop for access to all of the AI-enabled apps on a system. It’s similar to AMD’s game manager, but for AI-enabled applications, and it offers similar features, like the ability to track metrics from within the applications.
AMD Install Manager is another new addition. AMD Install manager is basically a tool that gives users the ability to manage AMD-specific software. It offers auto-updating, uninstallation options, product information, etc. and is supported on all products that still work with AMD Adrenaline Edition software.
Next is AMD Chat. AMD Chat is GPU-accelerated, local, offline Chat Bot with image generation capabilities for Radeon RX 9000 graphics cards. In addition to traditional Chat Bot features, AMD Chat will be able to toggle Adrenalin Edition features on or off by simply asking the chat bot to do so, and it can help gamers understand what each setting or GPU feature does as well. It uses an LLM to generate text responses and help with prompt processing, including image generation, without having to explicitly ask the chat to ‘generate an image of…’. At this time, AMD Chat supports English and Simplified Chinese. It’s using Llama 3.1 8B for English and QN2 7B for Simplified Chinese. Image generation is based on Stable Diffusion XL 1.0, though that’s only available for the English version for now.
AMD will also be using AI to increase driver quality. The company is now automating some parts of the game testing process, using an in-house created AI model. The AI-enhanced tools are capable of detecting crashes, in addition to numerous visual anomalies, like shader or shape artifacts, square patches, popping or missing textures, stutters, discoloration and various other corruption or rendering issues.
For end users, the company is also introducing AMD Image Inspector. AMD Image Inspector leverages AMD’s custom AI model and is designed to drive game quality improvements, by detecting corruption in gameplay. Users must opt-in to use the feature, and if they do, it captures and sends diagnostic data in the form of text and images back to AMD for more detailed diagnostics and (hopefully) quicker fixes. The text information is comprised of diagnostic results generated by the model from a specified session, along with associated anonymous system information. Gameplay images are captured every 3 seconds when GPU utilization is below 90% and the game is running in full screen mode. The model then reports any instances of visual artifacts or instability to AMD, so the company can more quickly react to any potential driver issues.
Radeon RX 9070 XT And 9070 Expected Performance
AMD provided some expected performance data for both the Radeon RX 9070 and more powerful 9070 XT, relative to the previous generation Radeon RX 7900 GRE...
With the games, settings and resolutions used by AMD, the Radeon RX 9070 is shaping up to be about 20% faster than the Radeon RX 7900 GRE, on average -- in some games the uplift is a little higher and in others it is a bit less.
Using the same set of games (and settings, etc.), the Radeon RX 9070 XT's performance relative to the Radeon RX 7900 GRE is expected to be much better. The average uplift at 1440p looks be about 38%. At 4K things look even better, with about a 42% performance uplift. We don't suspect AMD would put out any shoddy numbers, but of course, take all vendor provided data with a grain of salt. It won't be long until independent reviews hit anyway.
AMD RDNA 4 And Radeon RX 9070 Series: The Wait Is Almost Over
All told, RDNA 4 looks to be a significant step forward for AMD. Although the company hasn’t produced a monster GPU to compete at the ultra-high-end of the market, AMD's Radeon RX 9070 and 9070 XT target the midsection of the PC gaming market with MSRPs below $600, where the volume is, and appear poised to make a big splash, assuming performance in the real world mirrors what AMD has disclosed today. Thankfully, we won’t have to wait much longer to find out. Radeon RX 9070 cards should be available on March 6, with news of a more affordable Radeon RX 9060 also coming soon.
Next week is going to be very busy around here. NVIDIA is launching the final, previously announced member of the RTX 50 series and AMD will unleash the 9070 and 9070 XT. We’re told availability of the Radeon RX 9070 and 9070 XT will also be good, so here’s hoping the next few weeks see some much needed stabilization in the consumer GPU market. Hopefully gamers that have been itching to upgrade can get their hands on a shiny, new GPU, at – or close to – MSRP.
