TL;DR: A used RX 7900 XTX averages $815 on eBay in August 2026 — the cheapest 24GB card you can buy, $450 under a used RTX 3090. For LLM inference it delivers 90–110% of a 3090’s decode speed. The discount exists because everything that isn’t llama.cpp-style inference — image generation, fine-tuning, CUDA-only tools — runs worse or not at all.

	Used RX 7900 XTX	Used RTX 3090	New Radeon AI PRO R9700
Best for	Cheapest 24GB LLM inference	24GB + full CUDA ecosystem	32GB new-in-box, warranty
Price (Aug 2026)	~$815 used / $929 new	~$1,264 avg used	~$1,531 street
VRAM / bandwidth	24GB / 960 GB/s	24GB / 936 GB/s	32GB / 640 GB/s
The catch	No CUDA: weak diffusion, fine-tuning, day-one support	5-year-old card, price up ~30% since winter	Slower memory, ROCm/Vulkan same as XTX

Honest take: If your workload is chatting with and coding against local LLMs through Ollama, llama.cpp, or LM Studio on Linux, the used 7900 XTX is the best price-per-token deal of August 2026 — buy it and pocket the $450. If you also want Stable Diffusion, LoRA fine-tuning, or day-one support for every new model, that $450 is exactly what the CUDA ecosystem costs, and the 3090 is still the smarter buy.

Check what your target models need against your VRAM with the VRAM calculator before deciding — the whole argument below assumes you want the 24GB tier.

Why the 7900 XTX is suddenly the cheapest 24GB card

The DRAM crisis has repriced the GPU market almost purely by VRAM capacity. Used RTX 3090s climbed from ~$966 last winter to a $1,264 average in August 2026 ($1,201–$1,299 fair range, 366 listings). Used RTX 4090s sit at $2,268 for the same 24GB. Intel’s 32GB Arc Pro B70 jumped 30% to ~$1,268, and AMD’s own 32GB R9700 streets at ~$1,531.

The 7900 XTX missed the surge. Best Value GPU’s August 2026 tracker shows used cards at $814.99 on eBay and new ones at $929 on Amazon — under the $999 MSRP it launched at in December 2022. Gamers are moving on to RDNA4 and RTX 50-series cards, and the AI crowd that drove every NVIDIA 24GB card up ~30% mostly skips AMD. That gap is the whole trade.

The spec sheet says the discount is not about the silicon. Navi 31 carries 6,144 stream processors across 96 compute units, and — the number that matters for LLM decode — 24GB of GDDR6 on a 384-bit bus at 20 Gbps: 960 GB/s of bandwidth, per TechPowerUp. That is more than the used 3090’s 936 GB/s and 1.5× the R9700’s 640 GB/s.

Dollars per GB/s of bandwidth, computed from the prices above (our arithmetic):

Card (Aug 2026)	Price	Bandwidth	GB/s per dollar
Used RX 7900 XTX	$815	960 GB/s	1.18
Used RTX 3080 10GB	$285	760 GB/s	2.67
Used RTX 3090	$1,264	936 GB/s	0.74
Used RTX 4090	$2,268	1,008 GB/s	0.44
New R9700 32GB	$1,531	640 GB/s	0.42

The used RTX 3080 still wins raw bandwidth-per-dollar, but its 10GB ceiling locks it out of everything interesting. Among cards that can actually hold a 27B model, nothing touches the XTX.

What it actually does: measured tokens per second

Decode speed for local LLMs is memory-bandwidth-bound, and 960 GB/s shows up in the numbers. All figures below are community measurements, not vendor claims.

7B class — faster than a 3090. In llama.cpp issue #20934 (March 2026 builds, Ubuntu 24.04), a 7900 XTX generated 167–177 tok/s on Llama 7B Q4_0 under Vulkan (RADV) and 129–144 tok/s under ROCm. A separate dual-XTX bench repo on ROCm 7.1.1 (Arch Linux, build 34ce48d97) measured 119–139 tok/s per card on the same model, with prompt processing at 3,487–3,801 tok/s. Qwen2.5-Coder 7B Q4_K_M lands at 110–114 tok/s on ROCm. For reference, the used 3090 measures 111.74 tok/s on Llama 3 8B Q4_K_M in the XiongjieDai llama.cpp suite — different model sizes, so treat it as class-level parity, not a head-to-head.

27B class — the real reason to own 24GB. Real data from a CachyOS + Vulkan llama.cpp setup: Qwen3.6-27B Q5_K_XL runs at 27.4 tok/s generation with 562 tok/s prefill at 15K context, climbing to ~37 tok/s on a fresh chat with a short prompt. The 3090 baseline on the lighter Q4_K_M quant is 40.13 tok/s — so the XTX gives up roughly 10–25% on dense 27B decode while running a heavier quant. Both cards hold Qwen3.6-27B Q4_K_M (16.8GB) with headroom for 32K context; see the 24GB tier guide for the full model menu at this capacity.

70B class — same story as every 24GB card. One XTX can’t hold Llama 3.3 70B Q4_K_M (39.6GB); partial offload rules apply. Two XTXs can: the dual-card ROCm 7.1.1 bench ran the full 39.59 GiB model at 13.0–13.4 tok/s with prompt processing at 330–341 tok/s. That is in the same band as dual-3090 llama.cpp splits (7–10 tok/s on-site anchor) — and two used XTXs cost $1,630, less than a single 3090 pair at $2,528.

The backend quirk: Vulkan beats ROCm for chat

The same pattern we measured on RDNA4 holds on this card: llama.cpp’s Vulkan backend out-generates ROCm on gfx1100 — 167–177 vs 129–144 tok/s on the 7B test above, a ~25% gap that issue #20934 attributes to bursty GPU utilization under ROCm during token generation. The issue remains open with no fix as of March 2026 builds.

ROCm still wins prompt processing (4,000–4,400 tok/s on 7B), so the practical split is:

• Interactive chat and coding: Vulkan (-DGGML_VULKAN=ON in llama.cpp, or LM Studio’s Vulkan runtime — zero ROCm install required).
• Long-context prefill-heavy work (RAG, agents re-reading files): ROCm, via -DGGML_HIP=ON -DGPU_TARGETS=gfx1100.

Ollama’s ROCm binary auto-detects the card on Linux. Unlike the RX 6700/6600 generation, the 7900 XTX needs no HSA_OVERRIDE_GFX_VERSION workaround — gfx1100 has been a first-class ROCm target for years and keeps the full feature set (hipBLASLt, multi-GPU peer-to-peer) in ROCm 7.2. If Ollama falls back to CPU anyway, that’s a different failure — read the fix guide.

Verify the card is actually detected and doing the work:

$ rocminfo | grep -m1 gfx
  Name:                    gfx1100

$ ollama ps
NAME           ID           SIZE     PROCESSOR    UNTIL
qwen3.6:27b    5f3c...      20 GB    100% GPU     4 minutes from now

Anything other than 100% GPU in that PROCESSOR column means silent CPU spillover and single-digit tok/s. On Windows, ROCm for ML remains preview-grade — use LM Studio’s Vulkan runtime there and skip the ROCm install entirely, or set up Ubuntu properly.

A real problem you’ll hit: 100W at idle on multi-monitor

This is the card’s classic ownership gotcha, and it matters for an always-on AI server. With two or three high-refresh monitors attached, early drivers pinned the VRAM clock at 2,500 MHz and the card idled at ~100W — documented across multiple AMD community threads. On a 24/7 box, 100W of idle waste is ~$13/month at the August 2026 US average of 18.44¢/kWh.

The fix, in order:

1. Update Adrenalin. Recent driver testing by HWCooling measured the one-vs-two-monitor idle delta at just 2.3–3.2W — AMD largely fixed this in software.
2. Enable VRR/Adaptive Sync on displays that support it. Notebookcheck measured idle dropping to ~20W on a 4K 144Hz VRR panel.
3. Headless server? Non-issue — no displays, no bug. This is the natural deployment for a dedicated inference box anyway.

Under load, budget for the full 355W total board power (AMD spec): at 18.44¢/kWh that’s about 6.5¢ per hour of heavy inference, ~$3.93/month at 2 hours a day — full always-on cost math here. A quality 750W PSU handles one card; size up for two.

What the $450 discount actually costs you

Every dollar saved versus the 3090 comes out of the ecosystem, not the inference speed. Be honest with yourself about which of these you’ll want:

Image and video generation. NVIDIA soundly outperforms AMD in Stable Diffusion workloads — Puget Systems’ testing put even midrange GeForce cards ahead of RDNA3 flagships, and Tom’s Hardware’s 45-GPU SDXL suite shows the same ordering. The XTX works in ComfyUI on Linux ROCm, but if diffusion is a primary workload, the 3090 earns its premium.

Fine-tuning. QLoRA stacks (bitsandbytes, Unsloth) are CUDA-first. AMD support exists in places but is version-fragile. If you plan to train, buy NVIDIA.

Day-one everything. New quant formats (NVFP4), new runtimes, and new models land on CUDA first. Muse Glimmer’s dFlash drafter reached AMD via Vulkan weeks after CUDA. You will occasionally wait.

Resale confidence. The 3090’s price went up 30% in eight months because AI demand chases CUDA. The XTX’s flat price is great for buyers and a warning for sellers.

None of this touches plain LLM serving. If the box exists to run Ollama or llama.cpp behind Open WebUI or a coding agent, the CUDA tax buys you nothing.

Who should buy it — and who shouldn’t

Buy the used 7900 XTX if:

• You want the 27B–35B model class fully in VRAM for the least money, on Linux.
• You’re building a headless inference server and comfortable with one rocminfo sanity check.
• You want dual-card 70B capability at $1,630 — the cheapest 48GB path on the market right now.

Skip it if:

• Diffusion, video generation, or fine-tuning is on the menu → used RTX 3090 at $1,264.
• You want 32GB and a warranty → the R9700 at $1,299 MSRP when findable, with the same AMD software story.
• You’re Windows-only and unwilling to touch Vulkan settings → NVIDIA is friction-free there.
• You’d rather not own hardware at all → an A100 80GB rents for $1.39/hr on RunPod, and renting beats buying below ~15 hours/week of use.

Tested references in this article: ROCm 7.1.1–7.2, llama.cpp March–August 2026 builds, Ollama with ROCm auto-detect, prices checked August 24, 2026.

Recommended Gear

• RX 7900 XTX 24GB — the cheapest 24GB card of August 2026; buy used ~$815 or new $929
• RTX 3090 24GB — the CUDA-ecosystem alternative at ~$1,264 used
• AMD Radeon AI PRO R9700 32GB — new-in-box 32GB if capacity beats speed for you

FAQ

Is the RX 7900 XTX faster than an RTX 3090 for local LLMs? On paper, slightly — 960 vs 936 GB/s of bandwidth. In practice they’re the same class: the XTX measures 167–177 tok/s on 7B Q4_0 (Vulkan) vs the 3090’s ~112 tok/s on 8B Q4_K_M, and gives up 10–25% on dense 27B decode. Buy on price and ecosystem, not speed.

Do I need HSA_OVERRIDE_GFX_VERSION for the 7900 XTX? No. gfx1100 is natively supported through ROCm 7.2 with the full feature set. That workaround is for RX 6700/6600-class cards.

Does Ollama work on the 7900 XTX? Yes — Ollama’s ROCm build auto-detects it on Linux. Verify with ollama ps showing 100% GPU. On Windows, use LM Studio’s Vulkan runtime instead; ROCm-on-Windows remains preview-grade for ML.

Can two 7900 XTXs run a 70B model? Yes, fully in VRAM: 13.0–13.4 tok/s measured on Llama 3.1 70B Q4_K_M under ROCm 7.1.1 with llama.cpp layer splitting. Two used cards cost ~$1,630 — the cheapest resident-70B setup available in August 2026.

Why is it cheaper than every NVIDIA 24GB card? AI demand chases CUDA, so NVIDIA 24GB cards absorbed the DRAM-crisis repricing and AMD’s didn’t. The discount is real and so is its cause: weaker diffusion and fine-tuning support, and slower day-one coverage of new models.

Sources

• RX 7900 XTX Price Tracker US, Aug 2026 — Best Value GPU
• AMD Radeon RX 7900 XTX Review (specs: 24GB, 960 GB/s, $999 MSRP) — TechPowerUp
• ROCm significantly lower token generation vs Vulkan on RX 7900 XTX (gfx1100), issue #20934 — ggml-org/llama.cpp
• RX 7900 XTX llama.cpp ROCm benchmark results (7B + dual-card 70B) — 1337hero/rx7900xtx-llama-bench-rocm
• Qwen3.6-27B Q5_K_XL on RX 7900 XTX, Vulkan llama.cpp measured numbers — @sudoingX
• ROCm 7.2 consumer GPU support matrix 2026 — CraftRigs
• AMD ROCm local LLM setup on RX 7900 XTX — Local AI Master
• RX 7900 XTX 100W idle on dual monitor — AMD Community
• Radeon multi-monitor idle power draw finally fixed — HWCooling
• RX 7900 XTX 20W idle with VRR enabled — Notebookcheck
• Stable Diffusion performance: NVIDIA GeForce vs AMD Radeon — Puget Systems
• Stable Diffusion benchmarks: 45 GPUs compared — Tom’s Hardware
• Used RTX 3090 price data (on-site anchor, ResalePrices Aug 17 2026) — RunAIHome
• US electricity rate 18.44¢/kWh Aug 2026 — ChooseEnergy

Last updated August 24, 2026. GPU prices move weekly — verify current listings before purchasing.