(Credit: AMD)
Following a teaser at CES 2025, AMD has at last fully revealed its next-generation Radeon RX 9070 and Radeon RX 9070 XT, calling to mind the once-famous ATI Radeon 9000-series graphics card branding. These new cards, priced at $549 and $599, respectively, lead the charge on AMD’s latest strategy to compete in the midrange graphics card market.
With claims of improved ray-tracing performance over the previous generation, as well as enhanced super-scaling and frame generation via FSR 4, these new cards show potential. Here's a look over AMD's new RDNA 4 graphics architecture and the first graphics cards in the Radeon RX 9000 lineup.
AMD Radeon RX 9000: Recalling a Golden Era in GPUs
AMD cleverly named its new graphics cards the Radeon RX 9000 series, paralleling an early family of ATI cards known as the Radeon 9000 series. These graphics cards landed way back in 2002 and marked a golden era for ATI, the creator and manufacturer of the Radeon series, before AMD acquired the company in 2006.
Back then, the original Radeon 9000 series dominated the graphics card market against Nvidia's GeForce 4 series and GeForce FX 5000 cards that replaced them. This time around, the Radeon RX 9000 series won’t try to stand up to Nvidia's top-end cards, as AMD has yielded the high end to the enormous Nvidia GeForce RTX 5090. Instead, AMD aims to conquer the space between the ultra-high-end GeForce RTX 5090 and the budget market. (The latter is typically dominated by last-generation graphics cards nowadays.)
RDNA 4 Improves Ray Tracing
AMD's new graphics architecture, RDNA 4, powers the Radeon RX 9000 series. It's an evolution of the RDNA 3 architecture that anchored the Radeon RX 7000-series graphics cards. The key focus of this new architecture, according to AMD, is on improving ray-tracing performance, so most of AMD's changes were made to the ray-tracing hardware.
(Credit: AMD)The new RDNA 4 ray accelerators have eight Ray/Box units and two Ray/Triangle units each, a twofold increase over the ray accelerators in RDNA 3. These processing units carry out the ray-tracing work and help determine which items in a scene the ray accelerators need to work on. AMD also added dedicated hardware for instance-transform functions and reduced the latency associated with what it refers to as "bounding volume box hierarchy" (BVH8) traversal operations. This will increase the number of in-scene objects that the ray-tracing units can analyze and prioritize at once, which should improve ray-accelerator performance.
(Credit: AMD)Another feature change, which AMD refers to as "oriented bounding boxes," further helps ray accelerator performance by reducing false-positive intersections. When determining what is touched by a light source, BVH bounding boxes that are traditionally axis-aligned are used. This creates blocky, stair-step patches that are processed by the ray accelerators; these would include areas of empty space, essentially wasting some of the hardware’s time. With oriented bounding boxes, this issue is significantly reduced, resulting in up to a 10% improvement in traversal performance.
(Credit: AMD)AMD also says it is improving its memory system architecture with RDNA 4. The queues for memory requests by resources were extended to allow more data to be in motion at once. AMD also shifted to an out-of-order access design, whereas RDNA 3 took memory requests strictly in order, creating delays and work stalls when a piece of data loaded slowly.
(Credit: AMD)This new approach overcomes that limitation by enabling other memory requests to be fulfilled while others are loading. This change will help memory-bandwidth usage throughout the GPU, but AMD says this should have a more pronounced impact on ray-tracing workloads.
FSR 4 Leverages More AI for Graphics Rendering
AI performance has been a massively hot topic in the graphics card industry, so AMD introduced more AI throughput boosts in its RDNA 4 architecture.
For one, AMD improved wave matrix multiply-accumulate (WMMA) performance and efficiency, and it has added support for new 8-bit floating point operations to equip its AI accelerators for new workloads and models. WMMA operations are processing instructions that enable GPU cores to process mathematical tensors in higher-precision formats. This unlocks more advanced AI operations and graphical features, with greater energy efficiency than standard shading hardware.
(Credit: AMD)The applications of AI are growing and changing constantly, but AMD’s Radeon RX 9000-series GPUs will leverage it to boost resolution-scaling and frame-generation performance in the latest version of its signature FSR graphics-enhancement technology, FSR 4. A key change in FSR 4 is the switch to using ML-powered upscaling, which runs on the AI accelerators inside of AMD Radeon graphics cards. Initially, this feature will only be available on Radeon RX 9000-series cards, but it will later be supported on previous generations of AMD graphics cards like the Radeon RX 7000 series.
(Credit: AMD)FSR 4 is similar to Nvidia’s new DLSS 4 technology and aims to do much the same thing. FSR 3 and FSR 3.1 used resolution super-scaling, which renders frames at a lower resolution and uses supporting processing units to artificially reproduce the scene at a sharper resolution. Like DLSS, FSR 3 also introduced frame generation to its mix, which uses supporting processing units to artificially generate a new frame and stitch it between two raw-rendered frames, thus increasing effective frame rates.
The fourth generation of FSR picks up where FSR 3 left off, working to generate even more artificial frames between each originally rendered frame to drive even higher frame rates, though this still comes with trade-offs in terms of image quality and latency. The new ML-powered FSR 4 will also reportedly help to sharpen image quality compared with FSR 3 and FSR 3.1, but we'll have to see how that works out in practice.
(Credit: AMD)Adding generated frames between rendered ones can lead to more perceived latency relative to the frame rate, however. To help with this greater latency, AMD is also introducing Anti-Lag 2, an updated variant of an existing AMD technology, a latency-reduction tech that might be key to enjoying high-speed gameplay under FSR 4.
At launch, AMD says it will have more than 30 FSR 4 titles ready, and that list is expected to grow to include more than 75 titles in 2025.
Detailing the Radeon RX 9070 and RX 9070 XT
AMD's first two cards in the Radeon RX 9000 series, the Radeon RX 9070 and the faster Radeon RX 9070 XT, use the same graphics chip, with a die area of 357mm2 that consists of 53.9 billion transistors. This makes the GPU die roughly similar to Nvidia’s GB203 GPU die that powers the Nvidia GeForce RTX 5080. (It covers 378mm2 and contains 45.6 billion transistors.) Both fabricate the chips using similar TSMC manufacturing processes, too, with the Nvidia GB203 created on TSMC’s 4N process and the new RX 9070 GPUs created on its N4P process node.
(Credit: AMD)In total, AMD’s new GPU die has 64 compute units with 64 shaders each, adding up to a total of 4,096. These are accompanied by 256 texture mapping units (TMUs), 128 raster operation processors (ROPs), 128 AI accelerators, and 64 ray accelerators. All of these resources are active and available on the Radeon RX 9070 XT, but eight of the 64 compute units are disabled on the Radeon RX 9070, reducing its core counts to 3,584 shaders, 224 TMUs, 112 AI accelerators, and 56 ray accelerators.
The number of ROPs does not change from the Radeon RX 9070 XT to the RX 9070. AMD also added 64MB of its Infinity Cache, which it has previously said drives up to 3.25 times the bandwidth of 256-bit 16Gbps GDDR6, increasing memory resources within the same memory interface. Both graphics cards work on this enhanced 256-bit-wide memory interface, connecting to 16GB of 20Gbps GDDR6 video memory.
Other than the difference in core counts, the most significant differences between the two cards come from clock speeds and power draw. The Radeon RX 9070 XT will be clocked at 2.97GHz and require 304 watts of power, while the RX 9070 will be clocked at 2.52GHz and require 220W.
The RX 9070 cards will be available via AMD's usual slate of board partners, including Acer, ASRock, Asus, Gigabyte, PowerColor, Sapphire, and XFX. Most cards will employ the usual 8-pin power connectors, and all should support DisplayPort 2.1a and HDMI 2.1b ports to power the latest high-refresh-rate gaming monitors.
Performance and Pricing for Radeon RX 9070 and 9070 XT: Launching March 6
Due to the new architectural changes, projecting the latest Radeon RX 9000-series cards' performance is difficult at this time. AMD's Radeon RX 9070 XT clearly has fewer hardware resources than the older AMD Radeon RX 7900 XTX, but this doesn’t necessarily indicate it will be slower. AMD says that, in ray-tracing workloads, the Radeon RX 9070 XT should be about 10% faster than the RX 7900 XTX, but we will need to test this in the lab before we can say for sure.
(Credit: AMD)The fact that improving ray tracing performance was a key focus for AMD in developing the new RDNA 4 architecture is promising, as that was the weakest link in the RDNA 3 architecture. However, less seems to have been done to the shaders, and AMD says that in games that don’t support ray tracing, we may see the older Radeon RX 7900 XTX hold the upper hand over the RX 9070. That, coupled with Nvidia's comparable cards (read: the GeForce RTX 5070) close on the horizon and prices in flux as AMD's new cards hit the market, will make for murky waters for customers buying their next GPU.
Though we don't know exactly how well these cards will perform quite yet, if AMD is able to ship enough of these new graphics cards at MSRP, or close to it, it may steal some of the GeForce RTX 50-series spotlight from Nvidia. Its prices put the Radeon RX 9070 XT and the Radeon RX 9070 in direct competition with the Nvidia GeForce RTX 5070. Nvidia has struggled to ship enough of its new RTX 50-series graphics cards to meet demand so far, with the already-released models constantly out of stock and selling for inflated prices. If AMD can avoid supply shortages of its own, and 9070 prices stay in check, it could get an upper hand in this middle part of the graphics card market while Nvidia works to improve its supply issues.