AMD's RX 9070 (XT) - Finally Released and Better Than Expected?
[12:36 Tue,4.March 2025 by Rudi Schmidts]
After a long back and forth, AMD has finally made its new RX 9070 graphics card generation available. And since the cards have allegedly been available at retailers for over a month, the question remains why AMD waited so long to actually launch the product. One of the reasons was definitely the pricing. Since the direct competitors from Nvidia (RTX 5070 / TI) were not yet available, AMD simply didn't know how to assess their performance. But since the end of last month, it is known that a) Nvidia's cards have not made any major leaps in the mainly relevant gaming performance and b) that Nvidia's cards are exceptionally poorly available and therefore extremely expensive.
And so AMD now does not have to call for extreme fighting prices for its new models at the launch, but can, as expected, enter the market only slightly below Nvidia. In the USA, the recommended retail prices are now for the RX 9070 and for the "top model" RX 9070 XT - both excluding taxes. In Germany, the RRP will probably start at 649 and 699 including taxes.
In terms of pure computing power, AMD's largest GPU in 2025 is thus quite competitively priced. With almost 50 FP32 TFlops, the RX 9070 XT is almost exactly on the level of the Nvidia RTX 5070 Ti, but the memory with 640 GB/s is about 40 percent slower than Nvidia's 896 GB/s. Especially in video editing, memory throughput plays an important role.
Those who are primarily interested in GPU compute tasks, however, will find an interesting detail at AMD: Unlike Nvidia, AMD can double its FP16 TFlops through register halving. So, those who use or train KI models mainly with FP16 quantization get even 97.3 FP16-TFLOPs, double that of Nvidia's 5070 Ti. However, currently hardly any video application can work with FP16 accuracy, which is why Resolve and co. do not benefit from this.
It looks more exciting again with KI-TFLOPS, i.e. the accelerated matrix multiplication, which forms the foundation for almost every KI kernel. However, it is not always clear how manufacturers ultimately measure a peak value. For example, the sparse implementation and the accumulation register width are very variable. Therefore, a meaningful comparison is actually only possible in a practical implementation. In the published values, Nvidia and AMD are at least on par: AMD speaks of 779 TOPs (Peak INT8) and 1557 TOPs (Peak INT4). Nvidia correspondingly names 703 (INT8) and 1406 (INT4) as values for the RTX 5070 Ti.
A slightly more relevant data point from AMD's presentation a few days ago: Here, AMD claims that mask tracking under Resolve with the RX 9070 XT is 34% faster than with the RX 7900 GRE, which we had in the test here.
If we had to choose solely on the basis of the specs on paper, we would tend to recommend an Nvidia RTX 5070 Ti at an RRP of 879. Especially because of the faster memory and the 4:2:2 10-bit hardware decoding support. AMD has also improved its video decoding in the new cards, but has not mentioned 4:2:2 10-bit decoding at all. However, currently no RTX 5070 Ti is available in German retail for less than 1,180, which significantly worsens the real price-performance ratio. It also remains unclear whether Nvidia prices will move back towards the RRP in the near future - or whether they will remain at this high price level for the next few months. It could also become exciting if AMD releases an RX 9070 XT with 32 GB (which we keep hearing). Because Nvidia currently has nothing to counter a cheap KI-GPU with 32 GB of memory under 1000. And for many local KI applications, 16 GB is already significantly too little today.
;)
;)