AMD's Radeon RX 9070 and 9070 XT are here - at least almost

Contents

As expected, AMD cut the second slice of the RX-9070 salami in the livestream today: After the rather vague technical announcement at CES at the beginning of January, pretty much all technical data is official as well as a manufacturer's own selection of gaming benchmarks compared to the GeForce RTX 5070 Ti. AMD also announced the planned pricing in the stream: the Radeon RX 9070 will cost USD 549 and its bigger sister, the RX 9070 XT, will be available for USD 599. Compared to Nvidia's RTX 5000 UVPs, these are really competitive prices. As always, the prices do not include taxes and European surcharges, so you can expect to pay around a quarter more in euros than in US dollars. And board partners and retailers are also independent of these proposals in their pricing. So there is still a little suspense for the planned sales launch on March 6, when AMD is sure to have placed enough cards with retailers.

New graphics chip Navi48

The engine for the Radeon RX 9070 (XT) is the new, monolithic Navi48 graphics chip, whose up to 64 compute units include improved matrix units for AI functions and which is supported by 16 GB of GDDR6 memory in both the RX 9070 and 9070 XT. PCI Express 5.0 is also on board for the first time for AMD. On paper, the Radeon RX 9070 gets by with 220 watts, while AMD allows the XT version 304 watts of power consumption. The chip is manufactured using TSMC's N4 technology and, with 53.9 billion transistors in 356.5 square millimeters, achieves virtually the same circuit density as the Graphics Compute Die in the Navi 31 (45.4 billion transistors in 304 mm²). As with the 256-bit Navi32 of the previous generation, the last-level cache, typically called Infinity Cache by AMD, holds 64 megabytes, while the level 2 cache per memory controller is doubled to 2 megabytes, thus increasing from six to eight megabytes in total. Further technical details such as the clock rates can be found in the table.

AMD's Radeon RX 9070 and 9070 XT in comparison
	Radeon RX 9070 XT	RTX 5070 TI	Radeon RX 9070	RTX 5070	RTX 4070 S
Architecture	RDNA4 / Navi 48	Blackwell GB203	RDNA4 / Navi 48	Blackwell GB205	Ada Lovelace AD104
Shader cores / shader multiprocessors	4096 / 64	8960 / 70	3584 / 56	6144 / 48	7168 / 56
Base / boost clock rate	k.A. / 2,97 GHz	2,30 / 2,45 GHz	k.A. / 2,52 GHz	2,16 / 2,51 GHz	1,98 / 2,48 GHz
Throughput shader cores (FP32)	48,7 TFLOPS	43,90 TOPS	36,1 TFLOPS	30,8 TFLOPS	35,5 TFLOPS
Throughput tensor cores	1557 TOPS	1406 TOPS	1156 TOPS	988 TOPS	568 TOPS
Storage connection / data rate	256 Bit / 20Gbps	256 Bit / 28 Gbps	256 Bit / 20Gbps	192 Bit / 28 Gbps	192 Bit / 21 Gbps
Storage quantity / type	16 GByte GDDR6	16 GByte GDDR7	16 GByte GDDR6	12 GByte GDDR7	12 GByte GDDR6X
Memory transfer rate	640 GByte/s	896 GByte/s	640 GByte/s	672 GByte/s	504 GByte/s
Display	3 x DP 2.1a (UHBR13.5), 1 x HDMI 2.1b	3 x DP 2.1b (UHBR20), 1 x HDMI 2.1b	3 x DP 2.1a (UHBR13.5), 1 x HDMI 2.1b	3 x DP 2.1b (UHBR20), 1 x HDMI 2.1b	3 x DP 1.4a, 1 x HDMI 2.1
PCIe-Version / -Lanes	5.0 / x16	5.0 / x16	5.0 / x16	5.0 / x16	4.0 / x16
TDP	304 watts	300 watts	220 watts	250 watts	220 watts
Market launch	06.03.25	20.02.25	06.03.25	März 2025	17.01.24
Price (street)	599 US dollars plus tax	879 € (1100 €)	549 US dollars plus tax	649 €	639 € (750 €)

AMD is cautious in pre-selected benchmarks and positions the Radeon RX 9070 and 9070 XT against its in-house predecessor card, the Radeon RX 7900 GRE. According to the AMD benchmarks, the Radeon RX 9070 is around 20 percent (UHD: +21 percent) faster than the 7900 GRE over a course of 30 games in WQHD resolution. However, the performance increase for titles with ray tracing tends to be higher at 11 to 34 percent (UHD: 12 to 38 percent). With the Radeon RX 9070 XT, the performance increase doubles to 38 (WQHD) to 42 percent (UHD), with a maximum of 66 percent possible in ray tracing titles. Based on this again, the new cards should also be within striking distance of the Radeon RX 7900 XT and XTX respectively and at least the RX 9070 XT should also be able to come within the radius of the GeForce RTX 5070 Ti. In addition, AMD claims a 38 percent increase for the RX 9070 compared to the RX 6800 XT and 26 percent compared to a GeForce RTX 3080 at 4K resolution and the highest settings. The RX 9070 XT is said to be 51 percent faster than the RX 6900XT and 26 percent faster than the GeForce RTX 3090.

And although they have switched from the typical AMD naming scheme to an Nvidia-like one with increments of ten, AMD itself is not specifying whether the new Radeon cards should compete against the RTX 5070 (Ti) or its predecessor, the RTX 4070 (Ti/Super). Even though the AMD presentation contains images of a reference card, there will be no reference models, at least in Germany, and the Radeon 9000 field will be left entirely to the partners.

Aggressive positioning

Even during the presentation, it becomes clear that AMD wants to exploit all the supposed weaknesses of the competition. Among other things, the company emphasizes that upgrades will have an easier time with the Radeon RX 9070, for example by using proven eight-pin cables for the power supply, while Nvidia relies on the new, still not entirely unproblematic 12V 2x6 cables with 16 pins. These are only available on fairly new power supply units, many users have to make do with adapters.

However, the overall package shows that AMD is not relying on old technology. This also includes FSR4, which will run exclusively on Radeon 9000 cards and their MMA units with FP8 support with machine learning-supported upscaling. In addition, there are better display engines with DisplayPort 2.1 UHBR13.5 and HDMI 2.1b, which should also be more economical, especially in dual-display configurations. AMD also claims to have improved the video units; they should now achieve better quality at low bit rates and now also support AV1 with B-frames.

AMD cites its own data with relish, according to which 85 percent of gamers bought a graphics card for less than 700 US dollars - and thus hits the current pricing of the GeForce RTX 50 precisely. Of the three announced models RTX 5090, 5080 and 5070 Ti, there are currently de facto none below 1000 euros, although at least the latter card should actually be available for 869 euros.

The Radeon RX 9070 XT represents AMD's spearhead for gamers; the company will not be releasing a high-end graphics card for the time being. Accordingly, the Radeon RX 7900 XTX remains the fastest in-house model on paper in the overview.

RDNA4 - more ray tracing, more AI

The Radeon RX 9070 also marks the introduction of a new RDNA expansion stage, probably the last before UDNA. It is based on the large predecessor chip Navi31 with RDNA3 architecture. The Navi48 inherits the increased number of registers from it and these can now also be dynamically assigned to the wavefronts. This ensures fewer bottlenecks and therefore better utilization, not only, but especially when using ray tracing.

The ray tracing units have also been revised. In addition to twice as many intersection point calculations for bounding boxes and triangles, i.e. a doubling of the raw performance, they use memory more sparingly thanks to new data formats (BVH8 instead of 4). They can now run through the acceleration structure (BVH) with hardware support. Previously, this required normal shader operations, which were then missing for other calculations. In addition, memory requests are now possible "out-of-order", i.e. they no longer have to be processed in the order in which they were made – This also reduces the processing time for many memory-intensive calculations.

There is also a major leap in the "matrix accelerators", i.e. the acceleration functions for matrix multiply accumulates, which are central to most AI calculations. In RDNA4, FP8 and bF8 are additional 8-bit data formats. There is either a comparatively large range (E5M2) or greater accuracy (E3M4) of values to choose from. Unlike in RDNA3, the throughput in RDNA4 doubles from 16 bits (FP16, bF16) if 8-bit data formats are used. And the same is true for the step to 4-bit precision, which is only available for integer formats. The 256 FP32 operations per compute unit thus become up to 4096 INT4 operations, which can be doubled again with sparse matrices using 4:2 sparsity.

AMD itself sees a doubling of AI performance per compute unit with RDNA4 in Stable Diffusion XL 1.5, an AI model for image generation, among other things. The ML-based upscaling component in AMD's FSR4 also makes use of the higher throughput of FP8 and is therefore exclusive to Radeon 9000 graphics cards. According to AMD, FSR4 + Image Generation should increase the frame rates in many current games by a factor of 2 to 3.5.

(csp)