AMD Radeon RX 7900 XT

AMD Radeon RX 7900 XT: Launched December 13, 2022 at a $899 MSRP, $100 under the RX 7900 XTX it shipped beside, as a cut-down Navi 31: 84 compute units (5,376 stream processors) against the XTX's 96 (6,144), at 315 W total board power (PCPer's launch review notes AMD raised this from an original 300 W figure). The unusual 20GB follows from AMD's chiplet layout rather than from a marketing choice: each of Navi 31's six memory cache dies carries a 64-bit GDDR6 interface and 16MB of Infinity Cache, and this card enables five of them. Five dies make a 320-bit bus, 80MB of cache, 20GB (4GB per die) and, at 20 Gbps, exactly 800 GB/s. TechRadar's launch review headlined the card as "sitting in its big brother's shadow", and for local LLMs that framing mostly holds: the XTX's extra $100 at launch bought 20% more VRAM and 20% more bandwidth. Street prices for a four-year-old card are noisy; an eBay-listings tracker (GPU Poet) put its lowest-listing average near $633 in September 2026, without saying whether the listings were new or used.

The strongest models that fit this card are 27B-class. At 8k context, Qwen 3.8 27B and Qwen 3.6 27B fit at Q3_K_M (15.15 GB total with KV cache, 43.4 tok/s), but Q4_K_M comes to 19.02 GB against roughly 19.0 GB of usable VRAM (this site holds back 5% of the card's 20GB for the driver), so the usual sweet-spot quant tips into a slower offload at 35.8 tok/s unless context stays near 6K or below. Real GGUF sizes vary a little by builder, so a Q4_K_M build this close to the line can land on either side of it. Gemma 4 31B fits at Q3_K_M (18.23 GB, 37.4 tok/s). 35B-class mixture-of-experts models such as Ornith 1.5 35B-A3B and Qwen 3.5 35B-A3B fit at Q2_K (15.12 GB) and decode at 110.4 tok/s, because only about 3B parameters are read per token. 70B-class dense models do not fit at any quant: Llama 3.3 70B needs 40.72 GB even at Q3_K_M and offloads at 1.4 tok/s.

AMD lists the RX 7900 XT (gfx1100, AMD's identifier for this GPU family) as supported by ROCm 7.14 on Linux, limited to Ubuntu 24.04.4 and 22.04.5 or RHEL 10.1 and 9.7, and by the HIP SDK 7.2 on Windows 11; Ollama's GPU documentation lists the card on both operating systems. So the common advice that ROCm is Linux-only no longer applies to this card. The more useful choice is the llama.cpp backend. On the project's public llama-bench scoreboards (Llama 2 7B Q4_0; OS and driver not stated), this card generates 123.2 tok/s on Vulkan versus 116.2 on ROCm, while ROCm reads a 512-token prompt faster (3,098 versus 2,942 tok/s). A March 2026 llama.cpp issue on the same-silicon RX 7900 XTX (Ubuntu, Mesa's RADV driver) found llama.cpp's Vulkan backend generating about 20% faster than its HIP (ROCm) backend, with a maintainer replying that HIP isn't expected to beat Vulkan on this hardware, and a September 2026 comment still had Vulkan ahead on a 27B Q4 build (39.1 versus 36.0 tok/s). On Linux that makes Vulkan the easier first choice, with ROCm worth installing mainly when long prompts matter more than reply speed; none of this evidence is from Windows, so benchmark both backends there. Those scoreboard rows are also the best check on this site's estimates: its decode model (88% of rated bandwidth plus a fixed cost per layer) predicts about 136 tok/s for that 3.56 GiB file, so measured results land at roughly 85 to 91% of the estimate, and this page's tok/s figures are likely somewhat optimistic for small dense models. The estimate is also on the high side for mixture-of-experts models on AMD: one community report (llama.cpp discussion 15396, ROCm, August 2025) measured about 94 tok/s for GPT-OSS 20B fully on a 7900 XT, against this site's 121.2 and 129.5 tok/s for that model at Q5_K_M and Q4_K_M, because ROCm's expert-matmul kernels cost more per layer than CUDA's.