TL;DR: I want to try out running local AI - will 7900 XTX or 9070 XT give me the best experience?
What type of - and how big of a - difference will I experience running local AI on either 7900XTX (RND3) and/or 9070XT (RND4) on a Ryzen 9950x3d based machine?
As I understand it, RND4 support is still not ready, ie. a 9070xt will not be fully utilized and also “only” feature 16gb vram, where as the 7900 has 24gb.
So, it seems to me, that getting the latest generation of cards will give me less performance but that would seem a bit “backwards” to me.
Is there any expectations towards a releasedate for RND4 ?
Super depends on if you want to do LLM or Image Gen and if you’re on Windows or Linux.
If you’re on Linux 7900XTX is better for LLMs while 9070XT is better for Image Gen.
It’s the same on Windows, with the big difference being that for Windows the 9070XT is just kinda not supported at all right now. Whenever AMD gets around to releasing the rocm 6.4 sdk for Windows the support should get way better.
The 24GB of VRAM is actually a really big deal though. For LLMs this lets you comfortably run up to 32B. With just 16GB 14B is the biggest you can comfortably run and you could probably uncomfortably run 20B. Image Gen also benefits from the extra memory if you want to use models that use T5 like Flux (and you probably do because you can prompt in natural language and get good adherence with T5).
For local AI workloads, the RDNA 4 is the stronger pick it delivers faster AI compute, superior ray tracing, and better power efficiency thanks to its dedicated AI cores and FSR 4 support. Meanwhile, the 7900 XTX still leads in raw rasterization and offers more VRAM and bandwidth for large models
The benchmarks floating around comparing SDXL between the two cards are using the DirectML backend of ONNX with is quite a bit slower than the pytorch implementation. You can use ROCm on Linux or ZLUDA on Windows with Comfy and get almost 4x the IT/s on a 7900XTX. Theoretically the 9070XT **should** be faster with pytorch but at present ROCm and pytorch support for RDNA4 is buggy and not any faster than DirectM.
If SD performance is your main concern then you should probably hold off on it. If LLM performance is your main concern then the XTX is already just straight up better due to its better memory bandwidth.