The new RX 7900 XTX versus RTX 4080 SUPER local-AI comparison tackles a practical buying fork: 24GB of Radeon capacity or a 16GB Ada CUDA path.
The capacity difference appears before benchmarking
The official 19.9GB Qwen2.5-32B-Instruct Q4_K_M file is already 3.9GB larger than a 16GB pool before runtime overhead. A 24GB card leaves 4.1GB of raw file-size margin, though context, KV cache, runtime workspace and display use still count. The 15.9GB Q3_K_M file leaves an unsafe 0.1GB on 16GB versus 8.1GB on 24GB.
Software can reverse the practical choice
The Radeon path depends on the exact ROCm release, OS, framework and operator set. The RTX 4080 SUPER may be the more usable card for CUDA-dependent applications even with less memory. Neither specifications nor core counts prove speed. Use the CUDA versus ROCm checklist.
Compare the broader 16GB GPU paths and 24GB GPU paths, or run the hardware-fit calculator.
Build better evidence
Stage both exact benchmark starters and hold the model artifact, context, cache, batch, concurrency and timing constant. Record fallback, offload, exposed and peak memory, driver/runtime and wall power. Or watch both families privately without replacing saved preferences until you explicitly save.