AMD Radeon RX 9070 XT

What it does well

The RX 9070 XT is AMD's 2025 RDNA 4 mid-flagship at $700-900 — the strongest argument AMD has made for "skip CUDA" since the RX 7900 XTX. 16 GB GDDR6 at ~640 GB/s bandwidth is enough for 13B-class workloads at competitive decode speeds. ROCm 6+ matured through 2024 enough that day-to-day local AI on RDNA 4 + Linux works without the constant kernel-pinning dance the 7000 series demanded early. llama.cpp, Ollama, and vLLM all support RDNA 4 via ROCm at day-zero or close to it. Power draw at 304 W TDP is reasonable — single 8-pin + 6-pin connectors, 750 W PSU comfortable.

Where it breaks

• CUDA-locked stacks don't run. TensorRT-LLM, ExLlamaV2, SGLang — none have working ROCm paths or have them at quality parity with NVIDIA. If your workload, IDE plugin, or team deployment target is CUDA-first, the 9070 XT is fighting upstream.
• 16 GB caps daily-driver workloads at 13B-class. Same constraint as the RTX 5080. 32B-class needs 19-22 GB at Q4 — partial-offload territory. Pick the 7900 XTX (24 GB at similar pricing) if 32B-class is your goal.
• Day-zero new model support lags CUDA. ROCm wheels for new architectures land hours-to-weeks after the CUDA paths are working. For frontier-model users this matters; for "Llama 3.1 8B daily driver" users it doesn't.
• Windows ROCm is meaningfully worse than Linux ROCm. AMD's ROCm-on-Windows shipped in 2024 but remains second-tier. Linux is the production path; if you're on Windows + don't want WSL2, expect rougher edges.
• Resale value uncertainty. AMD GPU resale historically lags NVIDIA. RDNA 4 is new enough that the secondary-market floor isn't established yet. Plan to keep the card for its lifetime, not flip it.

Ideal model range

• Sweet spot: 13B-class at full 16-32K context — Qwen 2.5 14B, Phi 4 14B, R1 Distill Llama 8B at ~50-70 tok/s. Solid daily-driver capability.
• Stretch: 32B-class at Q4 with partial offload — drops to ~12-18 tok/s on most architectures. Functional for occasional use; not for daily.
• Comfortable: 7B-class at 100+ tok/s, embedding models, lightweight RAG pipelines, coding agents on smaller-tier models.
• AMD-specific advantage: works with Open WebUI and AnythingLLM just like NVIDIA cards do — the GUI tools layer is platform-agnostic.

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. 16 GB doesn't fit; partial-offload is single-digit tok/s. Pick 7900 XTX (24 GB) or NVIDIA 24-32 GB tier.
• Maximum tok/s on small models. A RTX 4070 Super at ~$600 used wins on $/throughput for sub-13B workloads in the CUDA ecosystem.
• Anyone Windows-first who doesn't want WSL2. Linux + ROCm is the production path. Windows works but feels like a port.
• Anyone who values predictable resale. AMD GPU resale is historically softer than NVIDIA. Don't buy this expecting 70%+ resale value at year 2.

Verdict

Buy this if you're Linux + ROCm-comfortable, your workload is 13B-class daily, you're price-sensitive enough that the $200-400 savings vs the 5080 matters, and you're not stuck in a CUDA-only stack. AMD's 2025 silicon + matured ROCm makes the 9070 XT the strongest "skip CUDA" pitch since the 7900 XTX — for the right operator.

Skip this if you need CUDA-ecosystem stacks (vLLM TP, TensorRT-LLM, EXL2, SGLang), if you're Windows-first, if your daily target is 32B-class or 70B (wrong VRAM tier), or if you value predictable software-update cadence over savings (NVIDIA's day-zero support is a real operator-time savings).

How it compares

• vs RTX 5080 (16 GB) → similar VRAM tier, NVIDIA wins on bandwidth + CUDA ecosystem maturity, AMD wins on price ($700-900 vs $1,100-1,300). For the 13B-class operator on Linux who's done with the CUDA tax, this is the operative comparison.
• vs RX 7900 XTX (24 GB) → 7900 XTX has 24 GB at ~$700-900 — same price tier, larger VRAM. Pick 7900 XTX for 32B-class headroom; pick 9070 XT for newer silicon (RDNA 4, better day-zero support, faster matured ROCm).
• vs RTX 4070 Ti Super (16 GB) → similar VRAM, different ecosystem. NVIDIA wins on CUDA + day-zero support; AMD wins on price-tier-relative if you're willing to take ROCm. The 9070 XT is the "stay AMD on new gen" play.
• vs Used RTX 3090 → 3090 used at $700-1000 with 24 GB VRAM beats the 9070 XT on $/VRAM and CUDA ecosystem. Pick 9070 XT only if you're committed AMD or want new-card warranty.
• vs RX 9070 (non-XT) → non-XT is ~$550-650 at similar 16 GB VRAM. The XT premium pays for ~15-25% more compute + clock headroom. Pick non-XT for value, XT if you want max RDNA 4 silicon.