Powered by the new RDNA 4 architecture and incorporating enhancements to rasterization, raytracing and machine learning, this may be AMD’s most significant GPU launch in years.
Powered by the new RDNA 4 architecture and incorporating enhancements to rasterization, raytracing and machine learning, this may be AMD’s most significant GPU launch in years.
Today AMD are formally unveiling the Radeon RX 9070-series of GPUs, the long-awaited next generation of products satisfying gaming requirements at 1440p while also incorporating some much-needed enhancements to their suite of hardware and software features. Encompassing two designs – the RX 9070 and flagship RX 9070 XT – these new cards will go toe-to-toe with NVIDIA’s GeForce RTX 5070-series where bang for your buck really matters.
The RX 9000-series was originally scheduled for a big reveal at CES 2025 in January but external factors, which we won’t speculate on here, gave pause to the higher-ups and caused an almost two month delay. AMD have spent this time wisely, optimising launch-day drivers and bolstering stock levels at retailers, so prospective owners will hopefully have a far more positive experience than other recent hardware releases.
With today’s livestream AMD are also finally announcing the price of the RX 9070-series cards. AIB pricing will be subject to whims of partners but the GPU architects have outlined an MSRP of $549 for the RX 9070 and $599 for the RX 9070 XT, significantly undercutting NVIDIA’s RTX 5070 Ti. Note also that this will not necessarily be reflective of any tariffs the current US administration has chosen to levy on hardware manufactured outside the country this coming quarter and so are subject to change.
RX 9070-series at-a-glance
AMD’s new GPUs are based on their latest RDNA 4 microarchitecture, which incorporates an array of targeted improvements over the prior generation. They both utilise the Navi48 GPU, silicon manufactured by TSMC on their 4N process for both performance and efficiency.
It is projected that the RX 9070-series will bridge the performance gap between the RX 7900GRE and RX 7900 XT, offering exceptional gaming experiences at 1440p (or 4K where appropriate). It is, according to AMD’s briefing, not expected to encroach on the position of the 7900XTX at the top of their product stack.
The RX 9070 XT boasts 64 CUs and 16GB of GDDR6 memory, clocking up to 2.97GHz within a 304W TBP envelope. This is a comparable number of CUs to the RX 7800 XT (60), underscoring perhaps the significant per-CU incremental improvements made in RDNA 4 over RDNA 3.
AMD project that the RX 9070 XT will have a performance premium of up-to 42% over the Radeon 7900GRE, the current best bang-for-your-buck card in their product stack. This improvement has been assessed over 30+ games and includes a mix of purely rasterised workloads and some which incorporate raytracing. Comparisons to the RTX 5070 Ti are preliminary and do indicate that it could be almost on par with NVIDIA’s new hardware, but time will tell on that score.
Sitting just below it with 54 CUs is the RX 9070. This card also boasts 16GB VRAM, offering a much larger frame buffer than many other cards in its class, while also clocking up-to 2.52GHz. Its AI processing capabilities will also be proportionally lower then the XT model, while still offering all the features. It will go head-to-head with the RTX 5070 at the same $549 price point when both cards are available next month.
With an anticipated performance improvement of 20% over the RX 7900GRE, the RX 9070 will an exciting card for 1440p gaming if the final retail price is right. It should offer a meaningful upgrade over an RX 6800 XT or RTX 3080 from 2020.
AMD’s target market are users on a 3-5 years upgrade cadence which need a little more horsepower than their current RX 6000-series or RTX 30-series GPU. They believe that the RX 9070-series is just the pair to tempt them into parting with older hardware and offer very healthy gains across a wide range of titles. What they don’t pretend to do is compete with NVIDIA’s high-end and halo silicon, leaving the RTX 5080 and 5090’s $1000 price bracket well enough alone.
As a final note, the RX 9070 XT utilises the whole of the Navi48 GPU. There is no scope for a higher-end SKU from this silicon in terms of CU count, though ‘XTX’ SKUs with enhanced operating frequencies are always possible.
Ray Tracing & AI
The wider adoption of raytracing in particular was identified as important to address in RDNA 4. AMD’s GPU architecture has typically been a generation or two behind their main competition, leading to an awkward capability miss-match as AAA titles integrated the feature further into game engines. In fact, this arguably came to a head late last year with Indiana Jones and the Great Circle, perhaps the first major release to out-right require hardware-accelerated ray tracing capabilities. While AMD’s last generation could just about handle these sorts of necessities, developers are clearly going to leverage the feature to an ever greater degree going forward.
With that in mind, RDNA 4’s 3rd Gen. Ray Tracing Accelerator adds a second ray tracing intersection engine which, when combined to other hardware and algorithmic improvements, effectively doubles the per-CU raytracing throughput at a lower VRAM cost. This will not double RDNA 4’s frame rates in raytraced games, but it will contribute towards better experiences long-term.
But that’s not the only factor involved in the performance of the latest games. AI upscaling has become something of a crutch to lean on as game engines try to simulate ever-more complex scenes, allowing frames to be rendered at a lower resolution than a display’s native resolution. Hardware accelerated AI tools have also become much prevalent in the creator and prosumer space, particularly image generation, manipulation and finalising.
While technically capable of accelerating Machine Learning workloads, RDNA 3’s 1st Gen AI accelerators were optimised for datacentres rather than incorporating aspects necessary for consumer-level features. RDNA 4’s 2nd Gen. Accelerators significantly improve throughput while also adding support for FP8 and sparse formats, increasing performance up-to 8x depending on datatype.
Software Features:
To build on the hardware improvements incorporated into RDNA 4, AMD software stack is also being augmented alongside the hardware release. This most notably arrives in the form of FSR 4, but also includes a one-click in-game settings optimiser and even noise AI cancelling for voice comms.
AMD FidelityFX Super Resolution 4 brings Machine Learning-enhanced upscaling to Radeon for the first time. This proprietary technology utilises the native FP8 support of RDNA 4 to process an ML model that upscales rendered frames to a monitor’s native resolution, similar to NVIDIA’s DLSS, and so is not supported on RDNA 3 or earlier hardware at this time. While ML upscaling had a rocky start, integration into the rendering pipeline and greater awareness of object transitions in a scene has vastly improved the quality of upscaled frames without significant impact on overall performance. FSR 4 should enhance frame rates and, in narrow situations, also improve visual fidelity.
In addition to upscaling, FSR 4 also optionally supports Frame Generation and Anti Lag features to enhance gameplay when the circumstances require it. FSR also technically supports multi-frame generation too; AMD however have reasoned that it provides no experiential benefit to a single inserted frame and the latency cost may not be worth it.
Titles which support FSR 3.1 should already be able to take advantage of FSR 4 through a common API and a single driver-level toggle. FSR 4 is also Neural Rendering ready, which upscales textures rather than full frames and is anticipated to greatly reduce the memory footprint of high-fidelity titles into the second half of the decade.
FSR 4 is expected to be supported by more than 30 games at launch and as many as 75 by the end of the year. Games that enable FSR4 with Frame Gen (inserting a single interpolated frame between rendered frames) can expect to enjoy performance up-to 3x faster than native in 4K gaming, jumping from 74fps to 222fps in CoD: Black Ops 6 for example.
The Right Price?
By coming in at $600 AMD have more than picked up the gauntlet NVIDIA threw down with the RTX 5070 Ti last month. However, while undercutting the latter’s MSRP by $150 sounds exceptionally bold, the key to its success will be the amount of stock available at retailers and how well AIBs adhere to AMD’s guidance on price.
NVIDIA’s launches this year have been plagued with issues beyond availability and price. This is an enormous opportunity for AMD but they also badly need a problem-free release for these cards in March. They’ve had a handful of weeks to build up stock at retailers they perhaps hadn’t initially accounted for, and more time to optimise day one drivers, so the final accounting will be solely on their shoulders.
The single un-forced error AMD seem to have inflicted is the RX 9070’s price. At $549 they’re only $50 less than the 9070 XT; close enough, perhaps, that they’re expecting 9070 customers to be up-sold to the more expensive part. It’s not the first time AMD have exploited the tactic but past success has always been mixed, so we wouldn’t be shocked to see a minor correction downwards next week. It may also come under pressure from the RTX 5070 presupposing NVIDIA’s card is more competitive than anticipated.
So now, the world waits on independent reviews next week. AMD are supporting factory TBP tweaks up to 340W at launch so performance and pricing might be a little more variable that it would seem at first blush. Expect to see partner cards from the likes of ASUS, Sapphire, Powercolor and XFX with both reference and factory overclocked specifications and their own idiosyncratic cooling solutions.
The Radeon RX 9070 and RX 9070XT launch on March 6th.
