AMD RX 9070 GRE AI, Blender benchmarks vs 9070 XT, 7800XT, Nvidia RTX 5070, 4070

Earlier this week, we shared the first part of our review of AMD's new RX 9070 GRE. It was about the gaming performance of the GPU, and we gave it an 8 out of 10. As a follow-up, similar to how we did with the 9070 XT and non-XT, we are doing a dedicated productivity review for the RX 9070 GRE as well, where we compare it against the 9070 XT, 9070, 7800 XT, as well as Nvidia's 5070 and 4070. This will include AI, rendering, compute, and more benchmarks.

AI performance, especially, is a very important metric in today's world, and AMD also promised big improvements thanks to its underlying architectural improvements.

We will be pitching it against the data we already have for the RX 9070, and RX 9070 XT, but also the Nvidia 5070 FE, MSI GeForce RTX 4070 VENTUS 2X 12G, and Gigabyte Radeon RX 7800 XT GAMING OC 16G as they are in a similar price class, but also because we do not have a comparable 5060 Ti card lying around here that we can compare it against.

Before we get underway, this is a collaboration between Sayan Sen and Steven Parker, who lent me his test bed. Also, there was no editorial input from AMD.

First up, the specs of the RX 9070, 9070 XT, and 9070 GRE, which were given to us by AMD:

As you can see from the specs above, it is less than the standard RX 9070 in every way that counts, except for slightly higher Boost and Game clock speed.

Design

Moving on, the RX 9070 GRE we were given is an XFX Swift triple-fan, dual-slot design with two 8-pin connectors. At 30cm (self-measured), it will fit in most systems easily. There is no RGB either.

The AMD Radeon RX 9070 GRE by XFX from all angles.

Test system

Our test system consists of the following:

• Lian Li O11 Dynamic Mini V2 Flow (Amazon|Newegg)
• ASUS Z890 ProArt Creator WiFi (Amazon|Newegg)
• Intel Core Ultra 7 270K Plus (Amazon|Newegg)

• Thermal Grizzly KryoSheet - 44x37 (Amazon|Newegg)

• 2x 16GB G.Skill Trident Z5 RGB (7200 MT/s in XMP) (Amazon|Newegg)
• Sabrent Rocket4 Plus 2TB SSD (Amazon)
• Windows 11 25H2 (Build 26200.8246)

AMD shared a press driver based on the recently released Adrenaline 26.5.2 that we were required to use. We now move on to our benchmarks. First up, we have Geekbench AI running on ONNX.

For some reason, the 9070 GRE does exceptionally well here in both half-precision (FP16) and single-precision (FP32). It manages to beat the RTX 5070 and RX 9070 non-XT, and is only behind the 9070 XT. Since Geekbench runs in short bursts instead of continuously hammering the graphics card, it seems the GRE's faster boost clocks are helping here.

Next up, we move to the UL Procyon AI test suite, starting with the image generation benchmark.

We chose the Stable Diffusion XL FP16 test since it is the most intense workload available on Procyon. The Nvidia cards do very well here, as even the 4070 out-muscles AMD's best fairy easily.

The positive thing about the GRE is that it gets quite close to the 9070 non-XT in this test; this indicates that the VRAM does not play a very big role here, as SD XL relies on float16 (FP16). So this is something to keep in mind again. If you wish to work with float32 AI workloads, graphics cards with larger than 12 GB buffers would likely emerge as victors. Regardless, the gains are still massive on AMD's 9000 series compared to the 7000 series.

Following image generation, we move to the text generation benchmark.

This is one test where the 9070 GRE struggled, quite a lot. It seems that the 12 GB VRAM and lower memory bandwidth of the new Radeon 9070 GRE are hurting it quite a bit; the split is massive, especially in a test like Llama2, which packs 13 billion parameters. As such, in all the tests, the 9070 GRE is the slowest of the lot.

Next, we tried Blender, and here the AMD GPUs were beaten by Nvidia. Rendering is something the Green team has always had a lead over the Red side, and it has not changed so far. On the positive side, though, the 9070 GRE shows significantly better results than the 7800 XT, which means AMD is on the right path. Catching up to Nvidia, though, will require a lot more effort. And we hope HIP and ROCm can keep improving.

Wrapping up AI testing, we measured OpenCL throughput in the Geekbench compute benchmark.

The RX 9070 GRE alongside the 9070 did not fare well here at all, even falling behind the 7800 XT. Interestingly, even the RTX 5070 could not beat the 4070 on OpenCL, so perhaps this suggests that OpenCL optimization may not have been a priority for either AMD or Nvidia in the modern era.

Conclusion

We reached the end of our productivity performance review of the 9070 GRE, and we have to say it's a mixed bag. Unlike the 9070 and 9070 XT, the GRE excels in some areas while losing ground fairly easily in others. Similar to how it happened in gaming, any time the card's memory subsystem gets hammered, it tends to fall behind the others. This was the case with text generation, wherein we saw the VRAM sometimes hit its maximum available 12 GB of usage with larger model sizes.

So what do we make of the RX 9070 as a productivity hardware? It can certainly be used, but you have to know it has its limitations.

For those looking for a GPU that can deal with more, AMD recently unveiled the Radeon AI PRO R9700, which is essentially a 32 GB refresh of the 9070 XT with some additional workstation-based optimizations. On a similar note, the new Ryzen AI Halo platform is something you can consider if you want to set up a local AI processing station.

Considering everything, we rate AMD's Radeon RX 9070 GRE a 7.5 out of 10 for its productivity performance. Price is less of a factor for those looking at productivity cases compared to those considering the GPU for gaming, and as such, we felt it did quite decently on many occasions and can be handy if you need a 12 GB GPU and, for some reason, don't want to get Nvidia.

Purchase links: RX 9070 / XT / GRE (Amazon US)

As an Amazon Associate, we earn from qualifying purchases.