AMD Radeon RX 7900 XT

What it does well

The 20 GB GDDR6 at ~800 GB/s bandwidth + $599-749 retail is the AMD answer to "I want more than 16 GB but the XTX premium isn't justified for my workload." Same RDNA 3 silicon as the RX 7900 XTX, 4 GB less VRAM, ~85% the compute, and ~$200-300 less. ROCm 6+ matured through 2024 to where consumer 7900-series is a real local-AI option for Linux operators. The 20 GB tier specifically: comfortably fits 13B-class FP16 (rare requirement), 32B Q4 with 16K context (common requirement), 70B Q3 fully on-GPU (single-digit tok/s but fits). 300 W TDP is reasonable — fits 750 W PSU comfortably.

Where it breaks

• 20 GB falls in an awkward middle. 4070 Ti Super at 16 GB is meaningfully cheaper for 13B-class workloads. 4090 used at 24 GB is the natural step-up if 32B-class is the goal. The 7900 XT's 20 GB sweet spot (32B Q4 with extra headroom) is real but narrow.
• CUDA-locked stacks don't run. TensorRT-LLM, ExLlamaV2, SGLang — none have working ROCm paths or have them at quality parity with NVIDIA. Production stacks like vLLM tensor-parallel work but trail CUDA.
• Day-zero new model support lags CUDA. ROCm wheels for new architectures land hours-to-weeks after CUDA paths are working.
• Windows ROCm second-tier vs Linux. Same caveat as the RX 7900 XTX and RX 9070 XT. Linux is the production path; Windows works but feels like a port.
• Resale floor is softer than NVIDIA equivalents. Plan to keep the card for its useful life, not flip it.
• The RDNA 3 generation is mid-2026 mature, not new. RDNA 4 (9070 series) is the newer silicon. Buying RDNA 3 in 2026 is a value play, not a "latest tech" play.

Ideal model range

• Sweet spot: 32B-class at Q4 with full 16K context — Qwen 3 32B, Qwen 2.5 Coder 32B, QwQ 32B at ~25-40 tok/s. The 20 GB unlocks this where a 16 GB 4070 Ti Super partial-offloads.
• Sweet spot (continued): 13B-class at full 32K context — Qwen 2.5 14B, Phi 4 14B at ~50-70 tok/s. Comfortable headroom.
• Stretch: 70B Q3 fully on-GPU (~30 GB partial-offload to system RAM) — single-digit tok/s, functional for occasional use.
• Comfortable: 7B-class at 90+ tok/s, embedding models, RAG pipelines, agent loops on small models.

Bad use cases

• Production CUDA stacks. vLLM tensor-parallel + Hopper FP8 + TensorRT-LLM ecosystem doesn't have an AMD answer at parity. Pick NVIDIA if your team's deployment target lives there.
• 70B daily-driver workloads. 20 GB is borderline; pick RX 7900 XTX (24 GB), RTX 4090 (24 GB used), or step up to 5090 (32 GB).
• Anyone Windows-first who doesn't want WSL2. Linux + ROCm is the production path.
• 13B-class only. 4070 Ti Super at $850-1000 with 16 GB GDDR6X is faster + cheaper for that workload tier.

Verdict

Buy this if you're Linux + ROCm-comfortable, your daily target is 32B-class at Q4 (where 20 GB's headroom over 16 GB matters), AND you specifically don't need the 7900 XTX's extra 4 GB. The 7900 XT is the right pick for "I want AMD + 20 GB at $200-300 less than the XTX" — a narrow but real operator preference.

Skip this if you need 24 GB (7900 XTX or 4090 used), if 13B-class is your ceiling (4070 Ti Super is faster + cheaper at the right tier), if CUDA is required, or if Windows is your primary OS without WSL2 acceptability.

How it compares

• vs RX 7900 XTX (24 GB) → XTX has 4 GB more VRAM + ~15% more compute at $200-300 more retail. Pick XTX for 70B headroom or maximum AMD perf; pick 7900 XT for the 32B-class sweet spot at a tighter budget. Both same RDNA 3 silicon, same ROCm story.
• vs RX 9070 XT (16 GB) → 9070 XT is newer RDNA 4 silicon at $700-900 with only 16 GB. 7900 XT wins on VRAM (20 vs 16); 9070 XT wins on newer silicon + faster day-zero ROCm support. Pick by VRAM-vs-newness preference.
• vs RTX 4070 Ti Super (16 GB) → similar pricing, NVIDIA wins on CUDA + ecosystem maturity, AMD wins on 4 GB more VRAM. For 32B-class workloads where the extra 4 GB unlocks full-GPU vs partial-offload, the 7900 XT wins on capability. For 13B-class, 4070 Ti Super wins on speed + ecosystem.
• vs Used RTX 3090 (24 GB) → 3090 used at $700-1000 has more VRAM (24 vs 20) + similar bandwidth + CUDA ecosystem maturity. Pick 3090 used if NVIDIA is acceptable and used-market is tolerable; pick 7900 XT new if you specifically want a new card with warranty or are committed to AMD.
• vs RTX 4080 Super (16 GB) → 4080 Super at $999 MSRP is faster on the workloads they overlap on but caps at 16 GB. 7900 XT wins on 32B-class capability; 4080 Super wins on 13B-class speed + CUDA.