The new RX 7900 XTX versus RTX 5090 local-AI comparison turns the 24GB-versus-32GB decision into exact model-file margins, software checks and a reproducible test plan.
The capacity boundary is real
For the official Qwen2.5-32B-Instruct Q4_K_M file, 24GB leaves 4.1GB before runtime overhead while 32GB leaves 12.1GB. The 26.9GB Q6_K file is larger than a 24GB card but leaves 5.1GB on a 32GB card before cache and workspace use. Neither result guarantees a fully resident run: context, KV cache, display use and backend allocations still count.
Software and power still decide the winner
The Radeon path requires an exact ROCm, OS, framework and operator check. The RTX 5090 benefits from CUDA but its Blackwell architecture also needs current driver and runtime support. Reference planning numbers differ sharply—355W board power and an 800W system recommendation for the RX 7900 XTX versus 575W and 1,000W for the RTX 5090 Founders Edition. Those specifications are not measured efficiency or speed.
Use the CUDA versus ROCm guide, compare other 32GB GPU paths, and run the fit calculator before buying.
Help build the evidence
If you can test either card, stage a matched benchmark with the exact model artifact, context, cache, OS, driver, backend, CUDA or ROCm version, unsupported operators, fallback, peak memory and wall power. Or watch both families for relevant listings.