AMD Radeon RX 7900 XTX vs AMD Radeon RX 7900 XT

Analysis

The RX 7900 XTX and RX 7900 XT launched together on December 13, 2022, at $999 and $899, and the difference that matters for local LLMs is how many of Navi 31's six memory cache dies each card enables. The XTX enables all six (24GB, 960 GB/s); the XT enables five (20GB, 800 GB/s), exactly five-sixths of the XTX on both capacity and bandwidth. Everything below follows from that one switch.

Speed follows bandwidth closely, with a little of the 20% gap lost to per-layer costs: Llama 3.1 8B at Q4_K_M decodes at 96.5 tok/s on the XT versus 111.6 on the XTX, and Qwen3 14B at Q4_K_M at 58.4 versus 68.2, a 16% to 17% edge for the XTX. Capacity matters more. Counting any quant, the two cards fit nearly the same models natively (51 versus 53 of 99, with Mixtral 8x7B and Nemotron 3.5 Lightning the only differences), because Q2 and Q3 builds squeeze into 20GB. At Q4_K_M, the quant most recommendations assume, the XTX fits 41 of the 96 models this site tracks with a standard quant ladder and the XT fits 32. Qwen 3.6 27B shows what that looks like: the XTX runs it at Q5_K_M (22.13 GB, 36.7 tok/s), while the XT's largest native build is Q3_K_M (15.15 GB, 43.4 tok/s, faster only because it moves fewer bytes per token). Q4_K_M, the step between them, comes to 19.02 GB, which fits the XTX at 41.8 tok/s but misses the XT's roughly 19.0 GB of usable VRAM by 0.02 GB and offloads at 35.8. Gemma 4 31B follows the same pattern: Q4_K_M on the XTX (22.63 GB, 36.3 tok/s) against Q3_K_M on the XT (18.23 GB, 37.4 tok/s). At 70B class neither card gets close: Llama 3.3 70B needs 40.72 GB even at Q3_K_M, so both offload, at 1.4 tok/s on the XT and 1.7 on the XTX. Software is a non-issue between the two: both are gfx1100, AMD lists both for ROCm on Linux and the HIP SDK on Windows, and llama.cpp results on this silicon have Vulkan generating tokens faster than ROCm.

Bottom line: The XTX is the better local-LLM card at launch pricing: 11% more money ($999 versus $899) bought 20% more VRAM and 20% more bandwidth, and the extra 4GB is what moves 27B to 31B models from Q3 up to Q4 and Q5. The XT makes sense when the street gap is wide. Spot checks of eBay-based trackers in late summer 2026 pointed at roughly $633 for the XT (GPU Poet, September) against roughly $815 for a used XTX (RunAIHome, August), about 29% more for 20% more VRAM, which leaves the XT slightly cheaper per GB but not by enough to settle it; those sources differ in method and month, so check current sold listings. Buy the XT if your models are 14B and under, 35B-class mixture-of-experts models at Q2 or Q3, or 27B-class models at Q3_K_M. Buy the XTX if you want Q4_K_M or better on the 27B to 31B class, or room for longer context. For 70B-class models, neither card is enough on its own.