TL;DR: As of October 20, 2026, a used RTX 3090 averages $1,533 — the same money as a new RTX 5080 ($1,546 tracked average). The 3090 buys 24GB and the whole 27B model class; the 5080 buys a warranty, GDDR7, and FP4, but hits a 16GB wall. A used RX 7900 XTX gets you the same 24GB from about $950 if you’ll run Linux.

	RTX 5080 (new)	Used RTX 3090	Used RX 7900 XTX
Best for	Fastest card for models ≤16GB, with warranty	24GB CUDA — the 27B class, long context	Same 24GB at the lowest price
Price (Oct 2026)	~$1,550–$1,800 in stock (MSRP $999)	$1,456–$1,558 fair range	from ~$950 used; $900–$1,358 new
The catch	16GB: Qwen3.6-27B and up won’t load	5–6 years old, zero warranty, mining risk	ROCm/Vulkan setup, weak image-gen path

Honest take: At price parity, buy VRAM, not a warranty — the used RTX 3090 is the pick for local AI. Only take the 5080 if everything you run fits in 16GB and you want FP4 image generation; take the XTX if saving ~$550 matters more than CUDA.

For three years the premise of the used RTX 3090 recommendation was that it undercut every new card with comparable bandwidth by $500 or more. That premise died this fall. ResalePrices’ fair asking range for a used 3090 hit $1,456–$1,558 across 320 tracked listings (October 20, 2026), while GPU Poet’s October tracking puts the cheapest RTX 5080 deals at a $1,546 average — with in-stock cards at major retailers mostly $1,799 and up (Walmart’s PNY listing sat at $1,799.99 in stock on October 1; Newegg’s Gigabyte AORUS was ~$1,900).

So the question is no longer “is the old card cheap enough to forgive its age.” It’s a straight trade at the same ~$1,500: 24GB of used GDDR6X, or 16GB of new GDDR7. And since AMD’s RX 7900 XTX quietly became the cheapest 24GB card on the used market, it belongs in the same decision. Run your own model and context numbers through the VRAM calculator as you read.

The spec sheet, and the one row that decides it

	RTX 5080	RTX 3090	RX 7900 XTX
VRAM	16GB GDDR7	24GB GDDR6X	24GB GDDR6
Memory bandwidth	960 GB/s	936 GB/s	960 GB/s
Board power	360W	350W	355W
Launched	Jan 2025	Sep 2020	Dec 2022
Warranty	Yes (AIB, typically 3 yr)	None (used)	None (used) / Yes (new)
Ecosystem	CUDA + FP4 (Blackwell)	CUDA	ROCm / Vulkan
Street price, Oct 2026	~$1,550–$1,800	$1,456–$1,558	~$950 used, $900–$1,358 new

Bandwidth — the number that actually sets token-generation speed on dense models — is a three-way tie: 960 vs 936 vs 960 GB/s. Power draw is a wash too: at $0.12/kWh all three cost roughly $0.042–$0.043 per hour under load (full power math).

That leaves exactly one row doing the work: VRAM. 16GB vs 24GB isn’t a 50% difference in practice — it’s a binary gate on which models load at all.

What each card actually runs, with measured numbers

These are measured results, not spec-sheet extrapolations. Sources are linked in each row’s citation and collected at the end.

Model class	RTX 5080	Used RTX 3090	Used RX 7900 XTX
7–8B Q4	132 tok/s (Llama 3 8B)	111.7 tok/s (Llama 3 8B Q4_K_M)	167–177 tok/s (Llama 7B Q4_0, Vulkan)
14B Q4	~58 tok/s	~45–50 tok/s class	~45–55 tok/s class
GPT-OSS 20B (MoE)	~160–185 tok/s (bandwidth-derived est.)	161 tok/s measured	runs, 24GB headroom for 32K ctx
Qwen3.6-27B dense Q4_K_M (16.8GB)	Spills — unusable (low-teens via CPU)	40.1 tok/s, fully resident	27–37 tok/s (Q5_K_XL, Vulkan)
Llama 3.3 70B Q4_K_M (42.5GB)	No	No (2–5 tok/s offloaded)	No (needs 2 cards: 13.0–13.4 tok/s)

Reading that table top to bottom:

On models that fit in 16GB, the 5080 is the fastest NVIDIA card here. 132 tok/s on the 8B class in LocalAIMaster’s testing beats the 3090’s 111.7 tok/s in the XiongjieDai llama.cpp suite by about 18% — GDDR7 plus three GPU generations of architecture doing what they should. The oddity is the XTX posting 167–177 tok/s under Vulkan in llama.cpp issue #20934 — but that’s a 7B model against the others’ 8B, so treat it as class-level parity-or-better, not a 30% win.

At 20B-MoE, the 3090’s number is measured and the 5080’s is an estimate. The 161 tok/s GPT-OSS 20B figure for the 3090 comes from the llama.cpp community benchmark thread #15396. The 5080 has no verified direct measurement we could find; scaling the RTX 4080’s measured 139.93 tok/s by the bandwidth difference lands at ~160–185 tok/s. Plausibly a tie, possibly a small 5080 win — but the 3090 runs this model at 16K–32K context with room to spare, where 16GB forces the context trap into play.

At 27B dense, the comparison ends. Qwen3.6-27B Q4_K_M is a 16.8GB file. A 5080 has roughly 15.5GB of usable budget after the OS and KV cache take their cut — the model does not fit, and what you get instead is partial CPU offload at low-teens tok/s. The 3090 runs it fully resident at 40.1 tok/s; the XTX runs the heavier Q5_K_XL quant at 27.4 tok/s at 15K context, climbing toward 37 tok/s on short prompts. Everything in the 24GB tier guide — Qwen3.6-27B, Gemma 4 31B at tighter quants, 20B-class models at long context — is simply absent from the 16GB menu.

At 70B, nobody here wins alone. 42.5GB doesn’t fit 24GB either; a single 3090 or XTX decodes at single-digit tok/s with CPU offload (the 70B-on-24GB reality). The only path to resident 70B in this price neighborhood is a second XTX — the dual-card ROCm bench ran the full model at 13.0–13.4 tok/s, and two used XTXs still cost less than one 5080.

The trap each card sets for you

RTX 5080: the spill is silent. Ollama will happily “run” a 27B model on a 16GB card — it just quietly pages weights through system RAM and decodes at a tenth of the card’s speed. Before blaming the card, check where the model actually lives:

$ ollama ps
NAME            SIZE      PROCESSOR          UNTIL
qwen3.6:27b     19 GB     23%/77% CPU/GPU    4 minutes from now

Anything other than 100% GPU in that PROCESSOR column means the card’s 960 GB/s is idling while your DDR5 does the decoding — that’s the mechanism behind every “my 5080 gets 12 tok/s” forum post. The fix is not a flag: pick a model that fits, buy 24GB, or rent. (OLLAMA_FLASH_ATTENTION=1 and OLLAMA_KV_CACHE_TYPE=q8_0 buy real KV-cache room at the margins; the 5080 deep dive has the full fit table.)

Used RTX 3090: you’re buying a 5-year-old mining survivor with no warranty. The card launched in September 2020; many spent 2021–2022 hashing 24/7. The practical mitigations from our 3090 buying guide: buy from a returnable source, avoid open-box “tested” cards from sub-99%-feedback sellers, and budget a thermal pad replacement — GDDR6X memory junction temps are this card’s known weak point under sustained inference load.

RX 7900 XTX: the backend choice is the setup. On gfx1100 (full XTX breakdown here), llama.cpp’s Vulkan backend out-generates ROCm by ~25% (167–177 vs 129–144 tok/s on the 7B test), while ROCm wins prompt processing — issue #20934 remains open. On Windows, ROCm for ML is still preview-grade: use LM Studio’s Vulkan runtime or set up Ubuntu properly. And if image generation matters to you, the ecosystem gap is real: ComfyUI’s fastest path in 2026 is NVFP4 on RTX 50-series — Blackwell-only, which is the one workload where the 5080 beats both 24GB cards outright.

Why the 3090 got this expensive (and why waiting won’t help)

The convergence isn’t a fluke of one tracker. The memory supply crisis that repriced DDR5 by 4× also repriced every GPU with AI-relevant VRAM, and NVIDIA is shipping no new consumer GPUs in 2026 — no 24GB card below the $2,500+ used RTX 4090 is coming to relieve demand. The used 3090 is up from roughly $1,050 in June to $1,533 today because it is the cheapest CUDA ticket to 24GB, and everyone building a home inference box knows it. That’s also why “wait for a dip” is a worse plan for GPUs than for RAM: the 3090’s supply only shrinks.

What to actually buy

Prices as of October 2026, all taken from the comparison above:

Your situation	The card	Price	Where
Everything you run fits in 16GB, you want a warranty and FP4 image gen	RTX 5080 16GB	~$1,550–$1,800	Check price
You want the 27B class and long context in CUDA — the local AI pick	Used RTX 3090 24GB	$1,456–$1,558	Check price
Same 24GB for ~$550 less, and Linux/Vulkan doesn’t scare you	Used RX 7900 XTX 24GB	from ~$950 used	Check price
Undecided — test your workload on a 24GB card first	Rented RTX 3090, from $0.07/hr	pay per hour	Vast.ai

If your model list genuinely stops at 20B — a coding assistant on GPT-OSS 20B, 8B-class chat, Gemma 4 QAT — the 5080 is a defensible buy: it’s the fastest thing here for that menu, it’s new, and it’s the only card of the three with a future in FP4 image generation. The moment “maybe I’ll want Qwen3.6-27B” enters your head, the 16GB card is the wrong purchase at any price near a 24GB one. That’s the whole verdict, and at October’s prices it costs nothing to follow it.

FAQ

Is a new RTX 5080 faster than a used RTX 3090 for local LLMs?

On models that fit in 16GB, yes — about 18% faster on the 8B class (132 vs 111.7 tok/s measured). On anything larger than ~15GB of weights plus cache, the 5080 stops being a contender at all: Qwen3.6-27B Q4_K_M runs at 40.1 tok/s on the 3090 and spills to CPU on the 5080.

Why does a used RTX 3090 cost as much as a new RTX 5080 now?

Because 24GB is the scarce resource. The AI-driven memory shortage pushed every high-VRAM card up (the 3090 rose from ~$1,050 in June 2026 to a $1,456–$1,558 fair range in October), NVIDIA is releasing no new consumer GPUs in 2026, and the only 24GB alternatives are the $2,500+ used 4090 or an AMD card.

Is the RX 7900 XTX actually usable for local AI in 2026?

Yes, with a Linux-and-llama.cpp-shaped asterisk. Under Vulkan it matches or beats the NVIDIA cards on small models and runs the full 24GB menu; the costs are a ~25% ROCm generation penalty (use Vulkan), preview-grade Windows tooling, and a much weaker image-generation path than NVFP4 on RTX 50-series.

Can any of these run Llama 3.3 70B?

Not resident on a single card — 42.5GB of Q4_K_M weights against 16–24GB of VRAM means CPU offload at 2–5 tok/s. The cheapest resident-70B setup in this price class is two used RX 7900 XTXs (13.0–13.4 tok/s measured), which still cost less than one RTX 5080 at October street prices.

Recommended Gear

• RTX 5080 16GB — fastest 16GB card; the pick when your models fit and you want a warranty
• Used RTX 3090 24GB — the 24GB CUDA value pick at $1,456–$1,558
• Used RX 7900 XTX 24GB — the budget 24GB seat, from ~$950

Sources

• RTX 3090 used fair asking range, $1,456–$1,558 across 320 listings — ResalePrices
• NVIDIA GeForce RTX 5080 price tracking, October 2026 ($1,546 lowest average) — GPU Poet
• PNY RTX 5080 at Walmart, $1,799.99 in stock Oct 1–2, 2026 — Trackalacker
• Used RX 7900 XTX live listings, cheapest $953 — GPU Poet
• RX 7900 XTX used price history — Best Value GPU
• RTX 5090 vs 5080 for local AI, measured tok/s (132 tok/s 8B) — LocalAIMaster
• GPU benchmarks on LLM inference (RTX 3090: 111.74 tok/s, Llama 3 8B Q4_K_M) — XiongjieDai
• GPT-OSS benchmark thread (161 tok/s on RTX 3090) — llama.cpp Discussion #15396
• ROCm vs Vulkan token generation on RX 7900 XTX (gfx1100), issue #20934 — ggml-org/llama.cpp
• RX 7900 XTX llama.cpp ROCm benchmark (7B + dual-card 70B at 13.0–13.4 tok/s) — 1337hero
• Qwen3.6-27B Q5_K_XL on RX 7900 XTX, Vulkan measured numbers — @sudoingX
• Comparing LLM performance on a 16GB VRAM GPU (GPT-OSS 20B, 139.93 tok/s on RTX 4080) — Rost Glukhov
• GeForce RTX 5080 official specs (16GB GDDR7, 960 GB/s, 360W) — NVIDIA
• AMD Radeon RX 7900 XTX review (24GB, 960 GB/s, 355W, $999 MSRP) — TechPowerUp
• RTX 3090 specifications (24GB GDDR6X, 936 GB/s) — NVIDIA Newsroom

Last updated October 20, 2026. GPU prices in this market move weekly — verify current rates before purchasing. The RTX 5080 GPT-OSS 20B figure is a bandwidth-derived estimate from a measured RTX 4080 result, flagged as such above.

Get the numbers before you buy

New GPU and mini-PC benchmarks, VRAM thresholds, and price checks — sent when there's something worth acting on, not on a schedule. No spam, unsubscribe anytime.

Email address Subscribe