Used RTX A6000 vs RTX 5090: Fit a 70B Model Before Buying

Forty-two and a half gigabytes. That single number settles most of this purchase, and almost nothing else about either card does.

The exact Llama 3.3 70B Q4_K_M file is 42,520,398,816 bytes: 42.52 decimal GB, or about 39.60 GiB. The file alone exceeds one 32GB RTX 5090 or two 16GB cards. A 48GB RTX A6000 is the capacity candidate in this three-card comparison, but loading the weights is only the first part of fitting a running model.

For the smaller models in our published RTX 5080 dataset, there is already a measured starting point. That does not establish which of these cards is fastest on every model that fits.

The line, and which cards cross it

Memory and power figures come from the manufacturer spec tables on the NVIDIA RTX A6000, RTX 5090 and RTX 5080 pages; PNY publishes 768GB/s of bandwidth and a dual-slot active fansink for its A6000 board.

These are archived Newegg feed prices, not current offers or used-card valuations. They no longer support a present-day price ranking. On October 16, the linked WINDFORCE RTX 5080 page showed Newegg’s own offer out of stock; its displayed price is not an available buying price. Amazon showed the matching model with an offer sold and shipped by TC Prime Electronics. Compare the seller, condition and return terms on both retailers before buying.

File size is not the runtime memory budget

Do not subtract a decimal download size from a GPU label and call the result free VRAM. The file conversion above uses bytes divided by 1,073,741,824 for GiB. Then check the actual GPU memory reported by NVIDIA SMI, including free and reserved memory, and the allocations logged by your inference runtime.

The llama.cpp server options expose GPU layer placement, context size, KV-cache types and multi-GPU splitting. Those settings change the working set. Test your intended context and concurrent requests; a file that fits does not guarantee the session fits. A 32GB card can also be part of a CPU/GPU offload setup. That is a different configuration from keeping this entire file on one GPU, and this page has no matched benchmark for it.

16GB is already fast, which is the awkward part

Our RTX 5080 throughput dataset recorded gpt-oss 20B at MXFP4 decoding 178 to 190 tokens per second across two dated captures, with the whole model sitting in 13,381 to 14,067 MiB of the card’s 16GB. Qwen 2.5 14B at Q4_K_M decoded at 94 to 97 tokens per second. Every cell is the median of three runs, and the published CSV carries the driver version, the Ollama build, the model digests and the resident memory readings.

That result is the reason the 48GB card is not an automatic upgrade. A 20B model and a 14B model are already resident and already quick in those captures. The reason to consider more memory is a workload that exceeds that measured setup, not a price ratio from the August feed. The VRAM-tier guide maps the full ladder, and the VRAM calculator works out the context-sensitive version of the same question.

When 32GB is useful, and when it is not

The RTX 5090’s 32GB cannot hold this entire 39.60 GiB file. That conclusion is specific to this artifact and full GPU residency; it does not rule out smaller quantizations, partial CPU offload or other models.

A working set above the RTX 5080’s capacity but within the RTX 5090’s available memory is a different purchase. Compare the exact model, context and throughput you need. The 575W NVIDIA specification in the table is board power, not a measured inference draw or a whole-system electricity estimate. No matched RTX 5090 or A6000 throughput number was captured on this page.

Compare a 16GB card when your working set fits

Our measured gpt-oss 20B result remains 178 to 190 tokens per second. Use that dataset’s model, runtime and settings when comparing it with another machine. NVIDIA lists 360W total graphics power for the RTX 5080; this is not the power measured during that inference run. A used card’s warranty and a new card’s seller terms also need checking separately.

The links below identify the WINDFORCE model; the dated table is not a promise of its current price or stock.

I may earn a commission if you buy through these affiliate links. As an Amazon Associate I earn from qualifying purchases.

The August feed contained thirteen in-stock RTX 5080 rows between $1,399.99 and $1,999.99. That is historical supply evidence only; it does not establish availability today.

The one case for going after a used A6000

Buy 48GB when your resident working set lands above 32GB and comfortably below 48GB. That is a useful target for this comparison. Parameter count alone is not a fit test: model architecture, quantization and context change the requirement.

Use the runtime’s allocation report to set the context budget. File bytes, available GPU memory and the KV cache are separate quantities. Test a sustained session at the intended context before the return period expires, rather than assuming a fixed 5.48GB of headroom from the product label.

Three things the A6000 gives you beyond raw capacity: ECC memory, a manufacturer-listed 300W maximum board power, and a two-slot board in a market where the fast consumer cards have grown to three and four. NVIDIA also supports it on the RTX Enterprise driver branch rather than the Game Ready branch, which is the right trade for a machine that works and the wrong one for a machine that games.

What to make the seller put in writing

A used A6000 is a private-party or reseller transaction, and the coverage rides entirely on that paperwork. PNY publishes a three-year limited warranty on a new card, and NVIDIA’s warranty policy sends most claims to the reseller or board partner first. Neither of those facts follows a card to its second owner on its own. So before you pay, get:

1. the exact part number and a photo of the serial;
2. whether the card is used, refurbished, open-box, or pulled from a workstation;
3. the return window, and who pays return shipping;
4. the warranty provider and the remaining duration;
5. enough of that return window left to run a full 48GB memory test and an hour of sustained load.

The fifth one is the one people skip. Run it the day the card arrives.

Compare any used offer with a current, available new-card offer for the same part number and condition. The $5,490.00 entry above was a new PNY board in the August feed; it is not a current price ceiling. Include return shipping and warranty coverage in the comparison. These links identify the PNY board:

How this was checked

The capacity argument uses one exact model artifact rather than a rounded parameter count: the Hugging Face repository reports 42,520,398,816 bytes for Llama-3.3-70B-Instruct-Q4_K_M.gguf, which is 42.52GB. Memory, power and cooling figures were read off the current NVIDIA and PNY specification tables on September 7. Throughput comes from our own RTX 5080 dataset, captured on an RTX 5080 under Ollama with the method and raw CSV published alongside it — no A6000 was benchmarked here, which is why this page recommends the 48GB card on capacity and never on speed.

Prices are the lowest in-stock USD offer per exact model from the August 17 Newegg feed, restricted to actual video cards. No used price appears anywhere on this page: every used-A6000 listing checked on September 7 was expired, dead, or unreachable. That is itself the finding. That September observation cannot establish today’s supply or price. Treat a new listing as a new offer to verify.

The October 16 correction re-read the exact Hugging Face file metadata, PNY specifications and runtime documentation. It separates historical prices from present buying advice, and file capacity from a tested runtime configuration. The original first-party throughput measurements above are unchanged.