R9700 Performance

Curious if anyone can comment on how the R9700 card does with various models. I’d love to setup a Rhea 1080p>4K test. But is the r9700 better with starlight being that its 32gb vs the 16gb and 24gb 9070xt/7900xtx? Seems like AMD still is a long way away from Nvidia for Topaz support. Wishing I bought a 3090 when I saw they were cheaper.

For me, using an AMD GPU for Topaz is borderline suicide. It’s not just about raw VRAM or compute

NVIDIA has a massive advantage in model support, CUDA/TensorRT and AI acceleration. 32 GB of VRAM are useless if the model doesn’t support your GPU.

Historically this is quite true, now, the gap is becoming smaller. Still a big gap though, agreed. I hate the 12VHPWR connector. The rate of failure is tiny but too much for my blood. So are the prices of 4090s and 5090s. Wishing now though I bought a 3090 when I was seeing exga xc3 for $700.

All that said, with my 9070xt I get equivalent-to or better results using Proteus and Rhea than a comparable 5070/5070Ti when they were priced similarly. So the CUDA gap is kinda irrelevant for that use.

You’re conflating CUDA cores with Tensor Cores. The CUDA gap isn’t really the point here. NVIDIA’s advantage for AI workloads comes from the combination of Tensor Cores and the CUDA/cuDNN/TensorRT software stack.

If your 9070 XT matches a 5070/5070 Ti in Proteus and Rhea, great, but that’s model-specific optimization, not evidence that the AI acceleration gap is irrelevant.

And this matters even more with diffusion models: they are far more compute-intensive and heavily dependent on dense matrix operations, making Tensor Core acceleration much more relevant than with conventional AI enhancement models. That’s where the NVIDIA advantage really starts to show.