TL;DR: A used RTX A6000 is the only sane way to get 48GB of CUDA VRAM on one card in 2026 — $2,600–$3,800 used, and it runs Llama 3.3 70B Q4_K_M fully resident at ~14.6 tok/s with zero multi-GPU configuration. The catch: it’s slower than a $1,264 used RTX 3090 on anything that fits in 24GB.

	Used RTX A6000	2× used RTX 3090	Rent A6000 on RunPod
Best for	One-card 70B, cramped cases, a future 96GB NVLink pair	Best 48GB per dollar, faster small models	Occasional 70B jobs
Price / Cost	$2,600–$3,800 used (Aug 2026)	~$2,528 for the pair	~$0.21–$0.49/hr
The catch	768 GB/s — slower than one 3090 on ≤24GB models	~700W load, PCIe/BIOS config work, 4–6 slots	Nothing local, nothing private

Honest take: If the whole point of your build is a dense 70B that just works — one power cable, one PCIe slot, no tensor-split flags — the A6000 is worth the premium over dual 3090s. If you’re optimizing dollars per gigabyte or you mostly run 27B–35B models, it isn’t: the 3090 pair is cheaper and each card is individually faster.

The used RTX A6000 occupies a strange spot in the 2026 used market. It’s a six-year-old Ampere workstation card that still sells for roughly triple what a used RTX 3090 goes for, and the reason is one number: 48GB. No consumer GPU has ever shipped with that much VRAM, NVIDIA’s current 96GB flagship repriced to $16K+, and the new RTX PRO 5000 72GB costs $9,100. Enterprise Ampere fleets are aging out, supply is finally loosening, and the A6000 has become the default answer to a very specific question: what’s the cheapest single CUDA card that holds a 70B?

This is the workstation entry in our used-GPU series — the RTX 3090, RTX 4090, and Tesla P40 verdicts cover the consumer and salvage tiers.

What $2,600–$3,800 actually buys

The RTX A6000 is the workstation twin of the RTX 3090: same GA102 Ampere silicon, more of it enabled, and double the memory (NVIDIA’s spec sheet):

Spec	RTX A6000	Used RTX 3090 (for scale)
VRAM	48GB GDDR6 (ECC)	24GB GDDR6X
Memory bandwidth	768 GB/s	936 GB/s
CUDA cores	10,752	10,496
Board power	300W	350W
Size / cooling	2-slot blower	3-slot open-air (most models)
NVLink	Yes — 2 cards pool 96GB	Yes (2-slot bridge, 24+24)
Used price, Aug 2026	$2,600–$3,800	$1,201–$1,299 ($1,264 avg)

Pricing first, because the spread is real. GPUDojo’s August 2026 tracking puts used units at $2,600–$3,800 as expiring enterprise workstation contracts push supply onto eBay, with clean “excellent condition” cards commanding close to $4,000 and retail-refurb offers starting at $3,999. RigPrice pegs the going rate at $4,500 — that’s what you’ll pay a reseller with a warranty, not what patient eBay bidding costs. Call it ~$3,200–$3,500 for a realistic private-party buy in August 2026, and treat anything under $2,800 as either a bargain or a card that spent four years at 100% fan duty in a render farm. Ask for nvidia-smi -q output before paying — the same seller-screening logic from our used 3090 guide applies, minus the mining-firmware worries (A6000s lived in workstations, not mining racks).

Note the bandwidth line in that table, because it’s the entire case against this card. The A6000 uses plain GDDR6, not GDDR6X, so it moves memory 18% slower than the 3090 it otherwise outclasses. Local LLM decode speed is bandwidth-bound, not compute-bound — which means on any model that fits in 24GB, the cheaper card wins.

The one-card 70B: real numbers

The reason to buy an A6000 is a dense 70B at the quantization people actually want. Llama 3.3 70B at Q4_K_M is 42.5GB of weights — it fits in 48GB fully resident with room for an 8K context cache, no offloading, no second GPU.

The multi-GPU llama.cpp benchmark set from XiongjieDai (CUDA 12.1, RunPod instances — dated May 2024, but Ampere llama.cpp performance has been stable for years) measured a single A6000 at:

• Llama 3 70B Q4_K_M: 14.58 tok/s generation, fully resident
• Llama 3 8B Q4_K_M: 102.22 tok/s (the same suite’s 3090: 111.74, 4090: 127.74)

That 14.58 tok/s deserves a double-take, because dual RTX 3090s running the same 70B with layer split land at 7–10 tok/s in our 48GB tier guide. One slower card beats two faster ones on the big model. The reason: with --split-mode layer (llama.cpp’s default and usually fastest multi-GPU mode), layers execute sequentially across the pair — you get 48GB of capacity but roughly one card’s worth of effective bandwidth, plus PCIe transfer overhead at every split boundary. The A6000 pays neither tax. It just runs.

Loading it looks like this — the line to watch is the layer count:

$ ./llama-server -m Llama-3.3-70B-Instruct-Q4_K_M.gguf -ngl 99 -c 8192
...
load_tensors: offloaded 81/81 layers to GPU
llama_kv_cache: CUDA0 KV buffer size = 2560.00 MiB

81/81 on one device is the thing you cannot get from any consumer card. On a 24GB card the same model runs partially offloaded at 8–14 tok/s — competitive on paper, but with system-RAM spillover, cold-start reload pain, and no context headroom. And 14–15 tok/s is above reading speed: fine for chat and code review, workable as a local backend for coding agents like Cline if you’re patient, slow for long agentic loops.

What 48GB does not unlock is the frontier-MoE class — gpt-oss-120b needs a ~60GB floor, and GLM-5.3-Flash’s smallest usable quant is 93GB. The full what-fits list lives in the 48GB tier guide; check your exact model+context combination in the VRAM calculator.

The EPS power trap (read before you plug it in)

Here’s the problem that catches nearly every first-time workstation-card buyer. The A6000’s 8-pin socket is not a PCIe 8-pin — it’s an EPS-12V 8-pin, the connector your motherboard uses for CPU power (PNY’s own Pro Tip on exactly this). The keying is close enough that people force the wrong cable, and forum threads are full of builders staring at both connectors trying to work out which cable is safe.

The failure mode in practice: card installed, PCIe 8-pin from the PSU forced or half-seated into the EPS socket → system powers on, nvidia-smi shows nothing or the card falls off the bus under load. The fix is the adapter that shipped in the retail box — a dual PCIe 8-pin → EPS-12V 8-pin pigtail. Retail PNY cards include it; used pulls from workstation fleets very often don’t, because Dell/Lenovo/HP chassis wired EPS natively. If your listing doesn’t show one, budget for the adapter (they’re inexpensive — search “dual PCIe 8-pin to EPS-12V GPU adapter”) and count it as part of the purchase. (A spare PSU CPU/EPS cable from your PSU’s own modular kit also works — never another brand’s, as modular pinouts differ.)

Two smaller quirks while you’re checking the listing photos:

• It’s a blower. The A6000 exhausts out the case rear, which is exactly what you want in a cramped or multi-GPU chassis and means it needs none of the duct-building surgery a passive Tesla P40 demands. The cost is noise under sustained load — a render-farm whine, not a gaming-card hum.
• ECC is toggleable. Workstation fleets often ran it enabled; ECC on GDDR6 reserves a slice of capacity and bandwidth for parity. For LLM inference you want it off (nvidia-smi -e 0, then reboot) — check the current state with nvidia-smi -q | grep -A2 "Ecc Mode" on any card you buy.

The longevity question — this one passes

Every used-GPU verdict now has to answer the driver question, because the answer killed our enthusiasm for the P40. CUDA 13.x dropped Maxwell, Pascal, and Volta support; everything Turing (compute capability 7.5) and newer remains in the current CUDA 13.3 toolkit.

The A6000 is Ampere, sm_86 — same architecture as the RTX 3090 that half the local-AI world still runs. It is inside the support window with no announced exit, flash-attention kernels target it, and every quant format llama.cpp ships runs on it. A used A6000 bought in late 2026 has no visible expiration date. That’s a real difference from the sub-$300 salvage tier, and part of why the card holds value: a $3,500 GPU you can still run in 2030 amortizes very differently than a $260 one that goes static in mid-2028.

Power math is equally boring, in a good way: 300W board power flat out — 50W less than a single 3090 and less than half of a dual-3090 rig under load — works out to about $0.036/hour at the $0.12/kWh US average (our arithmetic; your rate varies). It runs on the 650–750W PSU you already own.

NVLink: the quiet 96GB upgrade path

The A6000 supports third-generation NVLink: two cards plus a PNY NVLink bridge present a pooled 96GB address space at 112.5 GB/s of inter-GPU bandwidth (NVIDIA datasheet) — versus ~32 GB/s over PCIe 4.0 for bridgeless multi-GPU setups.

Two things make this more than a spec-sheet footnote. First, NVIDIA removed NVLink from the RTX 6000 Ada and every workstation card since — it’s a key reason the older A6000 still holds demand; the Ampere generation is the end of the line for consumer-reachable VRAM pooling. Second, 96GB is the tier where the current frontier-MoE quants live (GLM-5.3-Flash at 93GB, gpt-oss-120b with real context). Buying one A6000 today keeps a second-card path open that no dual-3090 or 5090 build can match — 96GB pooled at ~$7,000 used versus $16K+ for the RTX PRO 6000 Blackwell. Slower, absolutely. Less than half the price.

Why not the 6000 Ada, the PRO 5000, or the 3090 pair?

Path to 48GB+	Cost (Aug 2026)	Bandwidth	The trade
Used RTX A6000	$2,600–$3,800	768 GB/s	Slowest per-GB, simplest build, NVLink path
2× used RTX 3090	~$2,528	936 GB/s per card	Cheapest 48GB; config work, heat, ~700W
Used RTX 6000 Ada	$4,700–$6,500	960 GB/s	25% more bandwidth, no NVLink, +$2K
RTX PRO 5000 72GB (new)	$9,100	1,344 GB/s	72GB and warranty at triple the price

The RTX 6000 Ada is the same 48GB with ~2.35× the FP32 compute and 960 GB/s of bandwidth — and a used floor of $4,700–$6,500 that buys you roughly 25% faster decode for 60–80% more money, while giving up the NVLink path. For inference (bandwidth-bound), that’s a bad trade; for fine-tuning workloads that lean on its FP8 tensor cores, it’s a defensible one — though at that budget renting beats owning for most training runs.

Dual 3090s remain the value answer, with every caveat our dual-3090 build guide documents: IOMMU groups, ACS overrides, a 1,200W PSU, and two three-slot cards’ worth of heat. You save ~$700–$1,000 and gain speed on every model under 24GB. You spend a weekend earning it.

And if the 70B is a sometimes-need rather than a daily driver, don’t buy anything: A6000s rent from $0.21/hour across the cloud marketplaces that price-compare 16+ providers, RunPod among them — and $3,500 of hardware money funds literally thousands of hours of occasional batch work. Self-hosting the result long-term is aifoss.dev’s beat; the rent-vs-buy math is ours.

Verdict: buy or skip?

Buy the used A6000 if:

• You want a dense 70B fully resident on one card, in one slot, on one power cable — at ~14.6 tok/s it beats a dual-3090 split on both speed and sanity
• Your case, PSU, or patience can’t absorb a two-card build
• You want the only consumer-reachable path to a pooled 96GB (second card + NVLink bridge later)
• You’re buying at the honest end of the range ($2,800–$3,500) from a seller who’ll show nvidia-smi output

Skip it if:

• Dollars-per-gigabyte is the metric — dual 3090s are cheaper and faster per card
• You mostly run 27B–35B models: a single $1,264 RTX 3090 serves that class faster for a third of the price
• You want interactive-agent speed on 70B-class models — that’s the RTX PRO 5000/6000 tier or a rented H100, not any Ampere card
• The listing is $4,000+ — at that price you’re one good eBay week from 6000 Ada money, and the calculus changes

FAQ

Is the RTX A6000 the same as the RTX 6000 Ada or the old Quadro RTX 6000? No — three different cards, one terrible naming scheme. Quadro RTX 6000 (2018, Turing, 24GB), RTX A6000 (2020, Ampere, 48GB — this article), RTX 6000 Ada (2022, Ada, 48GB, no NVLink). Listings mix them up constantly; check the photos for “A6000” silk-screened on the shroud and GDDR6 capacity in nvidia-smi.

How fast is a 70B on one A6000, really? ~14.6 tok/s for Llama 3 70B Q4_K_M fully resident in llama.cpp benchmarks — above reading speed, faster than a dual-3090 layer split (7–10 tok/s), far below the 30+ tok/s of modern MoE models at this size class. Sparse models like Qwen3.6-35B-A3B run dramatically faster on the same card.

Does it need special power or cooling? It needs the EPS-12V adapter covered above (dual PCIe 8-pin → EPS, included with retail cards, often missing from used pulls) and nothing else — 300W total, active blower cooling, standard 2-slot PCIe 4.0. It’s the rare workstation card that drops into a normal desktop unmodified.

A6000 vs A40 — the listings look identical? Same GA102 silicon and 48GB, but the A40 is the passive datacenter variant: no fan, no display outputs by default, needs server airflow or P40-style duct surgery. For a desktop build, the A6000’s blower is worth the price gap; the A40 only wins in a rack.

Will it still get CUDA support in 2028? By every current signal, yes. CUDA 13 cut Maxwell/Pascal/Volta; Ampere (sm_86) remains fully supported in CUDA 13.3 with no announced deprecation, and it shares an architecture with the massively deployed RTX 3090 — NVIDIA orphaning it any time soon is hard to imagine.

Recommended Gear

• NVIDIA RTX A6000 48GB — the one-card 48GB path
• Used RTX 3090 24GB — the value alternative, one or a pair
• PNY NVLink Bridge for RTX A6000 — the later 96GB upgrade

Sources

• NVIDIA RTX A6000 product page & specifications — NVIDIA
• NVIDIA RTX A6000 datasheet (NVLink 112.5 GB/s, 300W, blower) — NVIDIA/Proton
• RTX A6000 48GB Used Price, August 2026 — GPUDojo
• NVIDIA RTX A6000 used price today — RigPrice
• GPU-Benchmarks-on-LLM-Inference (A6000/3090/4090 llama.cpp numbers) — XiongjieDai, GitHub
• Pro Tip #23: RTX A6000 and EPS-12V 8-pin GPU power — PNY
• RTX A6000 power connectors: supplied dual PCIe adapter or 8-pin CPU? — Level1Techs Forums
• NVIDIA to drop CUDA support for Maxwell, Pascal, and Volta — Tom’s Hardware
• CUDA Toolkit 13.3 release notes (supported architectures) — NVIDIA
• RTX 6000 Ada vs RTX A6000 (used floor, NVLink removal) — ITCT
• RTX 6000 Ada vs RTX A6000 for content creation — Puget Systems
• Running 70B models locally: exact VRAM by quantization — InsiderLLM
• gpt-oss-120b memory requirements: 60GB floor — YingTu
• RTX 3090 used price & fair asking range — ResalePrices
• RTX A6000 cloud pricing across providers — ComputePrices

Last updated August 31, 2026. Used-GPU prices move weekly; verify current listings before buying.