Radeon 7900 XTX tempts local LLM builders

This is not a launch, but it is a useful snapshot of where local inference economics are heading: CUDA still buys convenience, but the premium is harder to defend when llama.cpp/Vulkan performance on AMD is good enough for interactive 24B-35B quantized models.

• –RX 7900 XTX has the same 24GB VRAM class as RTX 3090, which matters more than brand loyalty for fitting Q4/Q5 local models
• –Reported real-world figures vary heavily by backend, but community numbers often land around 25-40 tok/s for roughly 30B-class quantized models on a tuned 7900 XTX
• –RTX 3090 remains the safer pick for CUDA-first tools, diffusion, training experiments, and broader framework support
• –At 495 euros versus 850 euros, the AMD card's value case is strong unless the buyer needs NVIDIA's software ecosystem more than raw local inference throughput