What is the best used GPU for local LLMs?

For most people a used GeForce RTX 3090: 24 GB of VRAM and 936 GB/s, enough for 27B–32B-class models at Q4 (Qwen3.8-27B at an estimated 33–47 tok/s), with full CUDA support. A new card with 24 GB or more starts at $949 list (Intel Arc Pro B70 32GB), so the used card only makes sense below that. On a tighter budget, a used 12 GB RTX 3060 runs 8B–14B models.

Is the Tesla P40 still worth it for local LLMs?

It gives 24 GB for the price of an old datacenter card, but it is slow: 346 GB/s of bandwidth (about 37% of an RTX 3090), weak FP16 math so long prompts process slowly, a passive cooler that needs a fan shroud, no display output and a CPU-style 8-pin power plug. It works in llama.cpp. Est. 42–59 tok/s on Qwen3-8B at Q4.

Is the AMD Instinct MI50 good for LLMs?

On paper yes: 32 GB of HBM2 at 1024 GB/s. The catch is software: AMD's current ROCm releases no longer support its gfx906 chip (ROCm 5.7 was the last fully supported release), so you rely on community ROCm builds or llama.cpp's Vulkan backend. It is also passively cooled. Buy it only if you are comfortable tinkering.

Used RTX A6000 or two RTX 3090s?

Both give 48 GB. The NVIDIA RTX A6000 is one 2-slot, 300 W card with 768 GB/s; two 3090s are 700 W of board power across two 3-slot cards, but each has 936 GB/s, so they decode a bit faster. Neither fits Llama 3.3 70B at Q4 with room to spare (50.2 GB on our formula), so 48 GB is really for 32B models with long context or 70B at lower quants. New, the cheapest 48 GB+ card is the NVIDIA RTX Pro 6000 Blackwell 96GB at $8,565 list.

Why don't you show used prices?

We only publish prices we can source and date. Used prices change daily and vary by condition and seller, so we show each card's launch MSRP for reference and, for each VRAM size, the cheapest new card at list price. That is your ceiling: a used card is only a deal below it. Use the eBay button on each card to see current listings.

What should I check before buying a used GPU?

Ask for a photo of the card running a GPU stress test and its memory temperatures, check that the VRAM size matches the listing, look for bent fins, missing screws or fan noise, confirm your power supply has the right connectors and enough watts (see the planner), and measure your case. Prefer sellers with returns.

Best used GPU for local LLMs

Used and previous-generation cards are the cheapest way to get a lot of VRAM. Here are the 40 used-market GPUs we track (gaming, workstation and retired datacenter cards like the Tesla P40, V100 and Instinct MI50), ranked by VRAM, with bandwidth, estimated tokens per second and the catches that listings don't mention.

Used GPUs tracked: 40
VRAM range: 8–48 GB
Bandwidth range: 224–1024 GB/s

Short answer: VRAM decides which models fit, bandwidth decides speed. A used RTX 3090 (24 GB, 936 GB/s) is the safe default. Datacenter cards (P40, V100, MI50) give 24–32 GB cheaply but need cooling and software work.

About prices: we don't track used street prices, so none are shown. Each card lists its launch MSRP (what it cost new, not what it sells for now), and the price ceiling table shows the cheapest new card with the same VRAM at today's list price.

Our short list

12 GB entry card

GeForce RTX 3060 12GB

VRAM: 12 GB
Bandwidth: 360 GB/s
Run at Q4: 350 / 656

Qwen3-8B at Q4: est. 43–62 tok/s. Biggest at Q4: OLMo-2-1124-13B-Instruct. The usual cheap way into 12 GB on NVIDIA.

Launch MSRP $329 (not a used price, source)
Pay less than a new Intel Arc B580 ($249 list) for this to make sense.

Amazon eBay

16 GB on AMD

Radeon RX 6800

VRAM: 16 GB
Bandwidth: 512 GB/s
Run at Q4: 388 / 656

Qwen3-14B at Q4: est. 34–49 tok/s. Biggest at Q4: Ling-Coder-lite. RDNA 2: not on AMD’s official ROCm support list for Radeon cards, though many users run it; llama.cpp Vulkan works without ROCm.

Launch MSRP $579 (not a used price, source)
Pay less than a new Radeon RX 7600 XT ($329 list) for this to make sense.

Amazon eBay

The classic 24 GB pick

GeForce RTX 3090

VRAM: 24 GB
Bandwidth: 936 GB/s
Run at Q4: 423 / 656

Qwen3.8-27B at Q4: est. 33–47 tok/s. Biggest at Q4: ERNIE-4.5-VL-28B-A3B-PT. Memory chips sit on both sides of the board, so check memory temperature under load. 3-slot reference cooler.

Launch MSRP $1,499 (not a used price, source)
Pay less than a new Intel Arc Pro B70 32GB ($949 list) for this to make sense.

Amazon eBay

24 GB on the cheap, with work

NVIDIA Tesla P40 24GB

VRAM: 24 GB
Bandwidth: 346 GB/s
Run at Q4: 423 / 656

Qwen3.8-27B at Q4: est. 12–17 tok/s. Biggest at Q4: ERNIE-4.5-VL-28B-A3B-PT. Passive cooler (needs server airflow or a fan shroud), no display output, CPU-style 8-pin (EPS) power input. Pascal has very slow FP16 math, so prompt processing is slow; llama.cpp runs it.

No list price: sold through OEMs.
Pay less than a new Intel Arc Pro B70 32GB ($949 list) for this to make sense.

Amazon eBay

32 GB with lots of bandwidth, with work

AMD Radeon Instinct MI50 32GB

VRAM: 32 GB
Bandwidth: 1024 GB/s
Run at Q4: 504 / 656

Qwen3-32B at Q4: est. 31–44 tok/s. Biggest at Q4: IQuest-Coder-V1-40B-Loop-Instruct. Passive cooler (needs a fan shroud). gfx906 is not supported by current ROCm releases (last fully supported: ROCm 5.7); community ROCm builds and llama.cpp Vulkan are the usual routes.

No list price: sold through OEMs.
Pay less than a new Intel Arc Pro B70 32GB ($949 list) for this to make sense.

Amazon eBay

32 GB on NVIDIA, with work

NVIDIA Tesla V100 32GB (PCIe)

VRAM: 32 GB
Bandwidth: 900 GB/s
Run at Q4: 504 / 656

Qwen3-32B at Q4: est. 27–39 tok/s. Biggest at Q4: IQuest-Coder-V1-40B-Loop-Instruct. Passive cooler (needs server airflow or a fan shroud), no display output. Volta: no FlashAttention 2 support; llama.cpp runs it.

No list price: sold through OEMs.
Pay less than a new Intel Arc Pro B70 32GB ($949 list) for this to make sense.

Amazon eBay

48 GB in one card

NVIDIA RTX A6000

VRAM: 48 GB
Bandwidth: 768 GB/s
Run at Q4: 511 / 656

Qwen3-32B at Q4: est. 23–33 tok/s. Biggest at Q4: AI21-Jamba2-Mini. 48 GB in a 2-slot blower card at 300 W, so two fit in a normal desktop.

Launch MSRP $4,649 (not a used price, source)
Pay less than a new NVIDIA RTX Pro 6000 Blackwell 96GB ($8,565 list) for this to make sense.

Amazon eBay

We may earn a commission when you buy through these links, at no extra cost to you.

Price ceiling: the new card with the same VRAM

A used card is only worth it below the price of a new card that does the same job. This table shows the cheapest new, current GPU we track with at least that much VRAM, at list price (MSRP). New cards also come with a warranty and current driver support.

For each VRAM size, the used GPUs we track and the cheapest new GPU with at least that VRAM at list price
VRAM	Used options	Cheapest new card with that VRAM	New list price
12–15 GB	GeForce RTX 3080 Ti, GeForce RTX 3080 12GB, GeForce RTX 4070 Ti, Radeon RX 6750 XT, Radeon RX 6700 XT, GeForce RTX 3060 12GB, GeForce RTX 2060 12GB	Intel Arc B58012 GB · 456 GB/s	$249list price (MSRP), source, 2026-10-22
16–20 GB	GeForce RTX 4080, Radeon RX 6950 XT, Intel Arc A770 16GB, Radeon RX 6800, Radeon RX 6800 XT, Radeon RX 6900 XT, NVIDIA RTX A4000	Radeon RX 7600 XT16 GB · 288 GB/s	$329list price (MSRP), source, 2026-10-22
24 GB	GeForce RTX 3090 Ti, GeForce RTX 4090, Radeon RX 7900 XTX, GeForce RTX 3090, NVIDIA RTX A5000, NVIDIA Titan RTX, NVIDIA Tesla P40 24GB	Intel Arc Pro B70 32GB32 GB · 608 GB/s	$949list price (MSRP), source, 2026-10-29
32 GB	AMD Radeon Instinct MI50 32GB, AMD Radeon Instinct MI60 32GB, NVIDIA Tesla V100 32GB (PCIe)	Intel Arc Pro B70 32GB32 GB · 608 GB/s	$949list price (MSRP), source, 2026-10-29
48+ GB	NVIDIA RTX 6000 Ada 48GB, NVIDIA RTX A6000	NVIDIA RTX Pro 6000 Blackwell 96GB96 GB · 1792 GB/s	$8,565list price (MSRP), source, 2026-10-28

Every used GPU, by VRAM

Sorted by VRAM, then bandwidth. Board power is the maker's reference spec. Launch MSRP is for reference only (how we calculate).

Used-market GPUs ranked by VRAM with bandwidth, board power, launch MSRP and the biggest model each runs at Q4
GPU	VRAM	Bandwidth	Board power	Launch MSRP	Biggest model at Q4	Models that fit
NVIDIA RTX 6000 Ada 48GBWorkstation	48 GB	960 GB/s	300 W	$6,799launch MSRP 2022, source	AI21-Jamba2-Mini~37.0 GB at Q4	511 / 656
NVIDIA RTX A6000Workstation · Ampere · 48 GB in a 2-slot blower card at 300 W, so two fit in a normal desktop.	48 GB	768 GB/s	300 W	$4,649launch MSRP 2020, source	AI21-Jamba2-Mini~37.0 GB at Q4	511 / 656
AMD Radeon Instinct MI50 32GBDatacenter · GCN 5.1 (gfx906) · Passive cooler (needs a fan shroud). gfx906 is not supported by current ROCm releases (last fully supported: ROCm 5.7); community ROCm builds and llama.cpp Vulkan are the usual routes.	32 GB	1024 GB/s	300 W	No list price (sold through OEMs)	IQuest-Coder-V1-40B-Loop-Instruct~28.7 GB at Q4	504 / 656
AMD Radeon Instinct MI60 32GBDatacenter · GCN 5.1 (gfx906) · Passive cooler (needs a fan shroud). gfx906 is not supported by current ROCm releases (last fully supported: ROCm 5.7); community ROCm builds and llama.cpp Vulkan are the usual routes.	32 GB	1024 GB/s	300 W	No list price (sold through OEMs)	IQuest-Coder-V1-40B-Loop-Instruct~28.7 GB at Q4	504 / 656
NVIDIA Tesla V100 32GB (PCIe)Datacenter · Volta · Passive cooler (needs server airflow or a fan shroud), no display output. Volta: no FlashAttention 2 support; llama.cpp runs it.	32 GB	900 GB/s	250 W	No list price (sold through OEMs)	IQuest-Coder-V1-40B-Loop-Instruct~28.7 GB at Q4	504 / 656
GeForce RTX 3090 TiGaming · Ampere · Highest-bandwidth 24 GB card of its generation; 450 W board power.	24 GB	1008 GB/s	450 W	$1,999launch MSRP 2022, source	ERNIE-4.5-VL-28B-A3B-PT~21.5 GB at Q4	423 / 656
GeForce RTX 4090Gaming · Ada	24 GB	1008 GB/s	450 W	$1,599launch MSRP 2022, source	ERNIE-4.5-VL-28B-A3B-PT~21.5 GB at Q4	423 / 656
Radeon RX 7900 XTXGaming · RDNA 3	24 GB	960 GB/s	355 W	$999launch MSRP 2022, source	ERNIE-4.5-VL-28B-A3B-PT~21.5 GB at Q4	423 / 656
GeForce RTX 3090Gaming · Ampere · Memory chips sit on both sides of the board, so check memory temperature under load. 3-slot reference cooler.	24 GB	936 GB/s	350 W	$1,499launch MSRP 2020, source	ERNIE-4.5-VL-28B-A3B-PT~21.5 GB at Q4	423 / 656
NVIDIA RTX A5000Workstation · Ampere · 24 GB in a 2-slot blower card at 230 W.	24 GB	768 GB/s	230 W	$2,250launch MSRP 2021, source	ERNIE-4.5-VL-28B-A3B-PT~21.5 GB at Q4	423 / 656
NVIDIA Titan RTXWorkstation · Turing · Turing: no FlashAttention 2 support; llama.cpp runs it.	24 GB	672 GB/s	280 W	$2,499launch MSRP 2018, source	ERNIE-4.5-VL-28B-A3B-PT~21.5 GB at Q4	423 / 656
NVIDIA Tesla P40 24GBDatacenter · Pascal · Passive cooler (needs server airflow or a fan shroud), no display output, CPU-style 8-pin (EPS) power input. Pascal has very slow FP16 math, so prompt processing is slow; llama.cpp runs it.	24 GB	346 GB/s	250 W	No list price (sold through OEMs)	ERNIE-4.5-VL-28B-A3B-PT~21.5 GB at Q4	423 / 656
Radeon RX 7900 XTGaming · RDNA 3	20 GB	800 GB/s	315 W	$899launch MSRP 2022, source	Voxtral-Small-24B-2507~17.9 GB at Q4	404 / 656
GeForce RTX 4080Gaming · Ada	16 GB	717 GB/s	320 W	$1,199launch MSRP 2022, source	Ling-Coder-lite~12.7 GB at Q4	388 / 656
Radeon RX 6950 XTGaming · RDNA 2 · RDNA 2: not on AMD’s official ROCm support list for Radeon cards, though many users run it; llama.cpp Vulkan works without ROCm.	16 GB	576 GB/s	335 W	$1,099launch MSRP 2022, source	Ling-Coder-lite~12.7 GB at Q4	388 / 656
Intel Arc A770 16GBGaming · Alchemist · 16 GB; check your runtime supports Intel GPUs (llama.cpp SYCL or Vulkan).	16 GB	560 GB/s	225 W	$349launch MSRP 2022, source	Ling-Coder-lite~12.7 GB at Q4	388 / 656
Radeon RX 6800Gaming · RDNA 2 · RDNA 2: not on AMD’s official ROCm support list for Radeon cards, though many users run it; llama.cpp Vulkan works without ROCm.	16 GB	512 GB/s	250 W	$579launch MSRP 2020, source	Ling-Coder-lite~12.7 GB at Q4	388 / 656
Radeon RX 6800 XTGaming · RDNA 2 · RDNA 2: not on AMD’s official ROCm support list for Radeon cards, though many users run it; llama.cpp Vulkan works without ROCm.	16 GB	512 GB/s	300 W	$649launch MSRP 2020, source	Ling-Coder-lite~12.7 GB at Q4	388 / 656
Radeon RX 6900 XTGaming · RDNA 2 · RDNA 2: not on AMD’s official ROCm support list for Radeon cards, though many users run it; llama.cpp Vulkan works without ROCm.	16 GB	512 GB/s	300 W	$999launch MSRP 2020, source	Ling-Coder-lite~12.7 GB at Q4	388 / 656
NVIDIA RTX A4000Workstation · Ampere · 16 GB single-slot card at 140 W.	16 GB	448 GB/s	140 W	$1,000launch MSRP 2021, source	Ling-Coder-lite~12.7 GB at Q4	388 / 656
GeForce RTX 3080 TiGaming · Ampere	12 GB	912 GB/s	—	$1,199launch MSRP 2021, source	OLMo-2-1124-13B-Instruct~10.6 GB at Q4	350 / 656
GeForce RTX 3080 12GBGaming · Ampere	12 GB	912 GB/s	350 W	$799launch MSRP 2022, source	OLMo-2-1124-13B-Instruct~10.6 GB at Q4	350 / 656
GeForce RTX 4070 TiGaming · Ada	12 GB	504 GB/s	285 W	$799launch MSRP 2023, source	OLMo-2-1124-13B-Instruct~10.6 GB at Q4	350 / 656
Radeon RX 6750 XTGaming · RDNA 2 · RDNA 2: not on AMD’s official ROCm support list for Radeon cards, though many users run it; llama.cpp Vulkan works without ROCm.	12 GB	432 GB/s	250 W	$549launch MSRP 2022, source	OLMo-2-1124-13B-Instruct~10.6 GB at Q4	350 / 656
Radeon RX 6700 XTGaming · RDNA 2 · RDNA 2: not on AMD’s official ROCm support list for Radeon cards, though many users run it; llama.cpp Vulkan works without ROCm.	12 GB	384 GB/s	230 W	$479launch MSRP 2021, source	OLMo-2-1124-13B-Instruct~10.6 GB at Q4	350 / 656
GeForce RTX 3060 12GBGaming · Ampere · The usual cheap way into 12 GB on NVIDIA.	12 GB	360 GB/s	170 W	$329launch MSRP 2021, source	OLMo-2-1124-13B-Instruct~10.6 GB at Q4	350 / 656
GeForce RTX 2060 12GBGaming · Turing · Turing: no FlashAttention 2 support; llama.cpp runs it.	12 GB	336 GB/s	184 W	No list price (sold through OEMs)	OLMo-2-1124-13B-Instruct~10.6 GB at Q4	350 / 656
GeForce RTX 2080 TiGaming · Turing · Turing: no FlashAttention 2 support; llama.cpp runs it.	11 GB	616 GB/s	250 W	$999launch MSRP 2018, source	Mistral-Nemo-Instruct-2407~9.5 GB at Q4	349 / 656
GeForce RTX 3080 10GBGaming · Ampere	10 GB	760 GB/s	320 W	$699launch MSRP 2020, source	HyperCLOVAX-SEED-Omni-8B~8.5 GB at Q4	340 / 656
GeForce RTX 3070 TiGaming · Ampere	8 GB	608 GB/s	290 W	$599launch MSRP 2021, source	NVIDIA-Nemotron-Nano-9B-v2~7.2 GB at Q4	315 / 656
Intel Arc A750Gaming · Alchemist	8 GB	512 GB/s	225 W	$289launch MSRP 2022, source	NVIDIA-Nemotron-Nano-9B-v2~7.2 GB at Q4	315 / 656
GeForce RTX 2080 SuperGaming · Turing · Turing: no FlashAttention 2 support; llama.cpp runs it.	8 GB	496 GB/s	250 W	$699launch MSRP 2019, source	NVIDIA-Nemotron-Nano-9B-v2~7.2 GB at Q4	315 / 656
GeForce RTX 3060 TiGaming · Ampere	8 GB	448 GB/s	200 W	$399launch MSRP 2020, source	NVIDIA-Nemotron-Nano-9B-v2~7.2 GB at Q4	315 / 656
GeForce RTX 3070Gaming · Ampere	8 GB	448 GB/s	220 W	$499launch MSRP 2020, source	NVIDIA-Nemotron-Nano-9B-v2~7.2 GB at Q4	315 / 656
GeForce RTX 2060 SuperGaming · Turing · Turing: no FlashAttention 2 support; llama.cpp runs it.	8 GB	448 GB/s	175 W	$399launch MSRP 2019, source	NVIDIA-Nemotron-Nano-9B-v2~7.2 GB at Q4	315 / 656
GeForce RTX 2070Gaming · Turing · Turing: no FlashAttention 2 support; llama.cpp runs it.	8 GB	448 GB/s	175 W	$499launch MSRP 2018, source	NVIDIA-Nemotron-Nano-9B-v2~7.2 GB at Q4	315 / 656
GeForce RTX 2070 SuperGaming · Turing · Turing: no FlashAttention 2 support; llama.cpp runs it.	8 GB	448 GB/s	215 W	$499launch MSRP 2019, source	NVIDIA-Nemotron-Nano-9B-v2~7.2 GB at Q4	315 / 656
GeForce RTX 2080Gaming · Turing · Turing: no FlashAttention 2 support; llama.cpp runs it.	8 GB	448 GB/s	215 W	$699launch MSRP 2018, source	NVIDIA-Nemotron-Nano-9B-v2~7.2 GB at Q4	315 / 656
GeForce RTX 4060 Ti 8GBGaming · Ada	8 GB	288 GB/s	160 W	$399launch MSRP 2023, source	NVIDIA-Nemotron-Nano-9B-v2~7.2 GB at Q4	315 / 656
GeForce RTX 3050 8GBGaming · Ampere	8 GB	224 GB/s	130 W	$249launch MSRP 2022, source	NVIDIA-Nemotron-Nano-9B-v2~7.2 GB at Q4	315 / 656

Which popular models run, and how fast

Cards with 12 GB or more. Estimated decode speed at Q4 in tokens per second (how we estimate speed); MoE models show an upper bound only. Older architectures may land at the low end of the range, and the P40's prompt processing is much slower than its decode speed suggests.

Estimated fit and decode speed at Q4 for popular models on each used GPU
Used GPU	Qwen3-8B ~6.7 GB	Qwen3-14B ~11.3 GB	gpt-oss-20b ~15.6 GB	Qwen3.8-27B ~20.4 GB	Qwen3-30B-A3B ~22.3 GB	Qwen3-32B ~23.8 GB
NVIDIA RTX 6000 Ada 48GB48 GB · 960 GB/s	Fits 116–164 tok/s	Fits 64–91 tok/s	Fits up to 374 tok/s	Fits 34–48 tok/s	Fits up to 449 tok/s	Fits 29–41 tok/s
NVIDIA RTX A600048 GB · 768 GB/s	Fits 93–131 tok/s	Fits 51–73 tok/s	Fits up to 299 tok/s	Fits 27–39 tok/s	Fits up to 359 tok/s	Fits 23–33 tok/s
AMD Radeon Instinct MI50 32GB32 GB · 1024 GB/s	Fits 124–175 tok/s	Fits 69–97 tok/s	Fits up to 399 tok/s	Fits 36–52 tok/s	Fits up to 479 tok/s	Fits 31–44 tok/s
AMD Radeon Instinct MI60 32GB32 GB · 1024 GB/s	Fits 124–175 tok/s	Fits 69–97 tok/s	Fits up to 399 tok/s	Fits 36–52 tok/s	Fits up to 479 tok/s	Fits 31–44 tok/s
NVIDIA Tesla V100 32GB (PCIe)32 GB · 900 GB/s	Fits 109–154 tok/s	Fits 60–85 tok/s	Fits up to 351 tok/s	Fits 32–45 tok/s	Fits up to 421 tok/s	Fits 27–39 tok/s
GeForce RTX 3090 Ti24 GB · 1008 GB/s	Fits 122–173 tok/s	Fits 68–96 tok/s	Fits up to 393 tok/s	Fits 36–51 tok/s	Tight up to 471 tok/s	Tight 30–43 tok/s
GeForce RTX 409024 GB · 1008 GB/s	Fits 122–173 tok/s	Fits 68–96 tok/s	Fits up to 393 tok/s	Fits 36–51 tok/s	Tight up to 471 tok/s	Tight 30–43 tok/s
Radeon RX 7900 XTX24 GB · 960 GB/s	Fits 116–164 tok/s	Fits 64–91 tok/s	Fits up to 374 tok/s	Fits 34–48 tok/s	Tight up to 449 tok/s	Tight 29–41 tok/s
GeForce RTX 309024 GB · 936 GB/s	Fits 113–160 tok/s	Fits 63–89 tok/s	Fits up to 365 tok/s	Fits 33–47 tok/s	Tight up to 437 tok/s	Tight 28–40 tok/s
NVIDIA RTX A500024 GB · 768 GB/s	Fits 93–131 tok/s	Fits 51–73 tok/s	Fits up to 299 tok/s	Fits 27–39 tok/s	Tight up to 359 tok/s	Tight 23–33 tok/s
NVIDIA Titan RTX24 GB · 672 GB/s	Fits 81–115 tok/s	Fits 45–64 tok/s	Fits up to 262 tok/s	Fits 24–34 tok/s	Tight up to 314 tok/s	Tight 20–29 tok/s
NVIDIA Tesla P40 24GB24 GB · 346 GB/s	Fits 42–59 tok/s	Fits 23–33 tok/s	Fits up to 135 tok/s	Fits 12–17 tok/s	Tight up to 162 tok/s	Tight 10–15 tok/s
Radeon RX 7900 XT20 GB · 800 GB/s	Fits 97–137 tok/s	Fits 54–76 tok/s	Fits up to 312 tok/s	Won’t fit	Won’t fit	Won’t fit
GeForce RTX 408016 GB · 717 GB/s	Fits 87–123 tok/s	Fits 48–68 tok/s	Tight up to 279 tok/s	Won’t fit	Won’t fit	Won’t fit
Radeon RX 6950 XT16 GB · 576 GB/s	Fits 70–99 tok/s	Fits 39–55 tok/s	Tight up to 224 tok/s	Won’t fit	Won’t fit	Won’t fit
Intel Arc A770 16GB16 GB · 560 GB/s	Fits 68–96 tok/s	Fits 38–53 tok/s	Tight up to 218 tok/s	Won’t fit	Won’t fit	Won’t fit
Radeon RX 680016 GB · 512 GB/s	Fits 62–88 tok/s	Fits 34–49 tok/s	Tight up to 199 tok/s	Won’t fit	Won’t fit	Won’t fit
Radeon RX 6800 XT16 GB · 512 GB/s	Fits 62–88 tok/s	Fits 34–49 tok/s	Tight up to 199 tok/s	Won’t fit	Won’t fit	Won’t fit
Radeon RX 6900 XT16 GB · 512 GB/s	Fits 62–88 tok/s	Fits 34–49 tok/s	Tight up to 199 tok/s	Won’t fit	Won’t fit	Won’t fit
NVIDIA RTX A400016 GB · 448 GB/s	Fits 54–77 tok/s	Fits 30–43 tok/s	Tight up to 174 tok/s	Won’t fit	Won’t fit	Won’t fit
GeForce RTX 3080 Ti12 GB · 912 GB/s	Fits 110–156 tok/s	Tight 61–87 tok/s	Won’t fit	Won’t fit	Won’t fit	Won’t fit
GeForce RTX 3080 12GB12 GB · 912 GB/s	Fits 110–156 tok/s	Tight 61–87 tok/s	Won’t fit	Won’t fit	Won’t fit	Won’t fit
GeForce RTX 4070 Ti12 GB · 504 GB/s	Fits 61–86 tok/s	Tight 34–48 tok/s	Won’t fit	Won’t fit	Won’t fit	Won’t fit
Radeon RX 6750 XT12 GB · 432 GB/s	Fits 52–74 tok/s	Tight 29–41 tok/s	Won’t fit	Won’t fit	Won’t fit	Won’t fit
Radeon RX 6700 XT12 GB · 384 GB/s	Fits 46–66 tok/s	Tight 26–36 tok/s	Won’t fit	Won’t fit	Won’t fit	Won’t fit
GeForce RTX 3060 12GB12 GB · 360 GB/s	Fits 43–62 tok/s	Tight 24–34 tok/s	Won’t fit	Won’t fit	Won’t fit	Won’t fit
GeForce RTX 2060 12GB12 GB · 336 GB/s	Fits 41–58 tok/s	Tight 23–32 tok/s	Won’t fit	Won’t fit	Won’t fit	Won’t fit

Before you buy used

Datacenter cards (Tesla P40, V100, Instinct MI50/MI60) have passive coolers built for server airflow: plan a fan shroud. Most have no display output, so you need another GPU or integrated graphics for the screen. The P40 takes a CPU-style 8-pin (EPS) power plug, so you may need an adapter.
Software support: FlashAttention 2 needs an Ampere or newer NVIDIA GPU (source); Pascal, Volta and Turing cards still run llama.cpp. AMD's gfx906 cards (MI50, MI60) are not supported by current ROCm releases (AMD's GPU table; ROCm 5.7 was the last fully supported version).
Condition: ask for a stress-test run with memory temperatures, check the VRAM size in the listing, and prefer sellers with returns.
Power and size: check board power, power connectors and card length against your PSU and case in the planner.
Two cards: two 24 GB cards give 48 GB. See multi-GPU for local LLMs and the dual used RTX 3090 and used RTX 3060 build recipes.

Head-to-head: RTX 3090 vs RTX 4090 · RTX 3090 vs RX 7900 XTX · RTX 4070 Ti Super vs RTX 3090. Have a card we don't list? Enter its VRAM in the VRAM calculator.

Quick answers

What is the best used GPU for local LLMs?: For most people a used GeForce RTX 3090: 24 GB of VRAM and 936 GB/s, enough for 27B–32B-class models at Q4 (Qwen3.8-27B at an estimated 33–47 tok/s), with full CUDA support. A new card with 24 GB or more starts at $949 list (Intel Arc Pro B70 32GB), so the used card only makes sense below that. On a tighter budget, a used 12 GB RTX 3060 runs 8B–14B models.
Is the Tesla P40 still worth it for local LLMs?: It gives 24 GB for the price of an old datacenter card, but it is slow: 346 GB/s of bandwidth (about 37% of an RTX 3090), weak FP16 math so long prompts process slowly, a passive cooler that needs a fan shroud, no display output and a CPU-style 8-pin power plug. It works in llama.cpp. Est. 42–59 tok/s on Qwen3-8B at Q4.
Is the AMD Instinct MI50 good for LLMs?: On paper yes: 32 GB of HBM2 at 1024 GB/s. The catch is software: AMD's current ROCm releases no longer support its gfx906 chip (ROCm 5.7 was the last fully supported release), so you rely on community ROCm builds or llama.cpp's Vulkan backend. It is also passively cooled. Buy it only if you are comfortable tinkering.
Used RTX A6000 or two RTX 3090s?: Both give 48 GB. The NVIDIA RTX A6000 is one 2-slot, 300 W card with 768 GB/s; two 3090s are 700 W of board power across two 3-slot cards, but each has 936 GB/s, so they decode a bit faster. Neither fits Llama 3.3 70B at Q4 with room to spare (50.2 GB on our formula), so 48 GB is really for 32B models with long context or 70B at lower quants. New, the cheapest 48 GB+ card is the NVIDIA RTX Pro 6000 Blackwell 96GB at $8,565 list.
Why don't you show used prices?: We only publish prices we can source and date. Used prices change daily and vary by condition and seller, so we show each card's launch MSRP for reference and, for each VRAM size, the cheapest new card at list price. That is your ceiling: a used card is only a deal below it. Use the eBay button on each card to see current listings.
What should I check before buying a used GPU?: Ask for a photo of the card running a GPU stress test and its memory temperatures, check that the VRAM size matches the listing, look for bent fins, missing screws or fan noise, confirm your power supply has the right connectors and enough watts (see the planner), and measure your case. Prefer sellers with returns.

Buying guides · Best GPU (new) · Best GPU by budget · Multi-GPU · GPU power · Tokens per second

Best used GPU for local LLMs

Our short list

GeForce RTX 3060 12GB

Radeon RX 6800

GeForce RTX 3090

NVIDIA Tesla P40 24GB

AMD Radeon Instinct MI50 32GB

NVIDIA Tesla V100 32GB (PCIe)

NVIDIA RTX A6000

Price ceiling: the new card with the same VRAM

Every used GPU, by VRAM

Which popular models run, and how fast

Before you buy used

Quick answers

New models, and what they need to run