When AMD announced its RDNA 4 desktop GPUs towards the end of last month, the company touted the glorious AI and ML performance improvements the new architecture packs, alongside gaming upgrades.
We tested the gaming performance of the RX 9070 XT in our dedicated piece. But since AMD is pretty proud of its AI enhancements, we are doing a separate small review for it to see how big of a boost the architectural improvements bring.
Before we get underway, this is a collaboration between Sayan Sen (author), and Steven Parker who lent us their test PC for this review. Speaking of which, here are the specs of the test PC:
- Cooler Master MasterBox NR200P MAX
- ASRock Z790 PG-ITX/TB4
- Intel Core i7-14700K with Thermal Grizzly Carbonaut Pad
- T-FORCE Delta RGB DDR5 (2x16GB) 7600MT/s CL36 (XMP Profile)
- 2TB Kingston Fury Renegade SSD
- Windows 11 24H2 (Build 26100.3194)
Drivers used for the 7800 XT and 9070 were Adrenaline v24.30.31.03 / 25.3.1 RC (press driver provided by AMD), and for the 4070 Geforce v572.47 was used.
We start with Geekbench AI and here the new 9070 XT was much better than the 7800 XT.
In half-precision (FP16), the RX 9070 XT was nearly 51% faster than the 7800 XT and 18% faster than the RTX 4070.
The single precision (FP32) gap isn't so big though as the 9070 XT beats the 7800 XT by 31%, meanwhile, against the 4070, the new AMD card sees a win by 24%.
UL Procyon's AI text generation is the next benchmark we ran and here, the 9070 XT is competitive against both the 4070 and the 7800 XT. This measures the text generation abilities of a hardware across four language models.
The Navi 48 GPU shows a 10-16% advantage over the 7800 XT though the gap between it and the 4070 is lower.
We also tried the UL Procyon Computer Vision test that measures machine vision inferencing performance. Computer vision is the ability of machines to extract and analyse information from images or videos.
We chose the FP32 precision for our testing since it is more resource-heavy than FP16. Unfortunately, the 9070 XT would crash every time we tried it indicating a compatibility issue. It is noteworthy here that both the RTX 4070 and RX 7800 XT worked fine in this test.
Finally, finishing up our AI testing, we have Procyon AI image generation. We chose Stable Diffusion XL FP16 for this as it is the most intense workload available in UL's current suite.
The RX 9070 XT shows the biggest win in this test over the 7800 XT as it comes out 83% faster and nearly catches up to the RTX 4070. The processing speeds of images for the GPUs are given below:
- RX 7800 XT: 36.289 secs/image
- RX 9070 XT: 19.819 secs/image
- RTX 4070: 18.285 secs/image
Seeing how it is the most intense test, we also noted the VRAM usage here and much like in the case of games, the Radeon GPUs did gobble up more memory, almost the entire available capacity.
Using HWiNFO (commercial license), the dedicated VRAM usage on the 7800 XT and 9070 XT were 15,980 MB and 15,735 MB respectively. The Nvidia RTX 4070 meanwhile was at 10,913 MB.
Wrapping up our AI testing there, we moved on to test productivity using Geekbench and Cinebench 2024.
In Geekbench computing using the OpenCL API, the 9070 XT underperforms both the 7800 XT and 4070.
There appears to be a bug in the OpenCL implementation at least in the case of Geekbench. AMD's own numbers put the compute throughput of the 9070 XT at up to 97.3 TFLOPS for FP16 while the 7800 XT is rated at 74 TFLOPS or a difference of nearly 31.5%.
We will update the charts with new scores if anything changes.
Cinebench 2024 was another test which did not work out as it did not detect our 9070 XT. To AMD's credit though, it did provide us later with a Blender patch that enables HIP (Heterogeneous-computing Interface for Portability) mode that is required for the benchmark to work on Radeon GPUs. We will update the data shortly after this review goes live as we are currently unable to do so due to personal inconvenience.
Conclusion
When we completed our game testing review of the RX 9070 XT, we were very impressed with how well-rounded the new AMD card was, finally; and so we expected a similar thing for AI and compute performance too.
All in all, we have to say AMD has delivered in this department too, but unlike the gaming part, this did not go flawlessly. There were problems here due to a couple of incompatibility issues but it is something we feel AMD can fix with driver updates and/or software updates. It is also catching up to and beating Nvidia's last-gen GPU which is another positive.
Outside of driver or software incompatibility problems, the only other thing that we can pick on the 9070 XT is the VRAM capacity. While 16 Gigs is plenty for games, for AI workloads, something like 20 GB would have been better.
We rated AMD's Radeon RX 9070 XT a 10 out of 10 in the gaming review. On this AI and compute performance review, considering all things, we think the card deserves slightly less, though, still a solid 8 out of 10 for its excellent performance.
Purchase links: RX 9070 / XT (Amazon US)
As an Amazon Associate we earn from qualifying purchases.
0 Comments - Add comment