Four RTX-era NVIDIA cards carry "6000" in their name: the Quadro RTX 6000 from 2018, the RTX A6000 from 2020, the RTX 6000 Ada Generation from 2022, and the RTX PRO 6000 Blackwell from 2025. Memory size is the quickest first check: 24GB, 48GB, 48GB and 96GB, in that order. The rule in one line: choose the RTX PRO 6000 Blackwell for a single-GPU workload that needs more than 48GB and fits in 96GB, and otherwise start with the RTX 6000 Ada Generation, unless you need NVLink.
Three older cards used the number too: the Quadro 6000, the Quadro M6000 and the Quadro P6000, all from before NVIDIA put "RTX" in the name. If a listing only says "nvidia quadro 6000", ask for the full model name before going further.
Provider pages add to the confusion. Nebius's RTX PRO 6000 Blackwell Server Edition page sits at a URL ending in "rtx-6000", and a Google Cloud blog post from 2 February 2026 reverses the word order to "RTX 6000 PRO" in its URL, although both use the full name in their text. CoreWeave's documentation says its eight-GPU instance was renamed from "RTX PRO 6000 Blackwell" to "RTX PRO 6000 Blackwell Server Edition (High Memory)". None of that is the reader's fault, and it is why the table below gives each card's full official name and a link to its rent page.
Which RTX 6000 you actually have
| Official name | Launched | Architecture | Memory | Name clues | Rent page |
|---|---|---|---|---|---|
| NVIDIA Quadro RTX 6000 | 2018 | Turing | 24GB GDDR6 | "RTX 6000," "Quadro RTX 6000" | Rent |
| NVIDIA RTX A6000 | 2020 | Ampere | 48GB GDDR6 | "A6000" | Rent |
| NVIDIA RTX 6000 Ada Generation | 2022 | Ada Lovelace | 48GB GDDR6 | "RTX 6000 Ada," mislabeled "RTX A6000 Ada" | Rent |
| NVIDIA RTX PRO 6000 Blackwell | 2025 | Blackwell | 96GB GDDR7 | "RTX PRO 6000," bare "rtx-6000" in URLs | Rent |
Launch years and architectures above come from the site's own spec table. NVIDIA's newsroom dates the Quadro RTX 6000 announcement at SIGGRAPH on 13 August 2018, the same event that introduced the Turing architecture. The RTX A6000 followed on 5 October 2020, per CGChannel and VideoCardz's coverage of the announcement, and ServeTheHome reports that the same release also retired the "Quadro" brand from NVIDIA's workstation GPUs entirely. Notebookcheck reports that NVIDIA initially called the third card simply "RTX 6000" before adding the "Ada Generation" suffix, reportedly to avoid confusion with the 2018 Quadro RTX 6000, the exact confusion this article exists to resolve. VideoCardz itself briefly ran a headline calling that same card the "RTX A6000 'Ada'" before the market settled on its real name, direct evidence that even trade press struggled with the transition. NVIDIA's newsroom dates the RTX PRO 6000 Blackwell family announcement at GTC on 18 March 2025, adding "PRO" as a new prefix across its whole Blackwell-generation workstation and server lineup.
The older cards that also answer to "6000"
NVIDIA announced the original Quadro 6000 at SIGGRAPH on 27 July 2010, part of its Fermi-based lineup with the Quadro 5000, 4000 and 2000. CGChannel and TechPowerUp describe a Quadro M6000 announced in March 2015 and a refresh with more memory on sale about a year later, which is why a secondhand "Quadro M6000" listing with no memory figure is still ambiguous today.
The Quadro P6000 followed on the Pascal architecture, with 24GB of GDDR5X memory at 432GB/s of bandwidth and a 250 watt board power, per the site's own spec table. NVIDIA's Quadro bridge guide pairs the Pascal-generation Quadro GP100 with NVLink and the P6000 with the older SLI HB bridge, and Puget Systems' bridge compatibility chart likewise places the M6000 before NVLink, so none of these three older cards offers NVLink. They suit legacy CAD or rendering machines better than new AI work. The P6000 still has a rent page if you need one for an old workload.
Full specs, read carefully
| Spec | Quadro RTX 6000 | RTX A6000 | RTX 6000 Ada | RTX PRO 6000 |
|---|---|---|---|---|
| VRAM | 24 GB | 48 GB | 48 GB | 96 GB |
| Memory type | GDDR6 | GDDR6 | GDDR6 | GDDR7 |
| Memory bandwidth | 672 GB/s | 768 GB/s | 960 GB/s | 1,792 GB/s |
| FP16 (dense) | 130.5 TFLOPS | 154.8 TFLOPS | 364 TFLOPS | 503.8 TFLOPS |
| FP16 (with sparsity) | Not published | 309.6 TFLOPS | 728 TFLOPS | 1,007.6 TFLOPS |
| FP8 (dense) | Not published | Not published | 728.5 TFLOPS | 1,007.6 TFLOPS |
| FP8 (with sparsity) | Not published | Not published | 1,457 TFLOPS | 2,015.2 TFLOPS |
| FP32 | 16.3 TFLOPS | 38.7 TFLOPS | 91.1 TFLOPS | 126 TFLOPS |
| TDP | 260 W | 300 W | 300 W | 600 W |
| Interconnect | NVLink, PCIe 3.0 | NVLink, PCIe 4.0 | PCIe 4.0 | PCIe 5.0 |
| Launch year | 2018 | 2020 | 2022 | 2025 |
| Architecture | Turing | Ampere | Ada Lovelace | Blackwell |
The table uses family-level specifications. For RTX PRO 6000 Blackwell, NVIDIA product pages and the source sheet flag differences by edition and clock mode, so do not assume every Server Edition matches the displayed throughput and bandwidth.
Precision support rewards a careful read. The spec table has FP8 figures only for the RTX 6000 Ada Generation and RTX PRO 6000 Blackwell. NVIDIA's TensorRT support matrix, dated 8 September 2026, lists FP8 support on Ada and Hopper examples but not its Turing or Ampere examples; it also notes that FP4 on Ada and Hopper is emulated, not accelerated. SemiAnalysis's architecture breakdown, published 23 June 2025, places native FP4 support at Blackwell. In the site spec table, RTX PRO 6000 Blackwell is the only one of these four families with native FP4 throughput figures.
Interconnect is the other trap, because NVLink support does not increase with each generation here. The site spec table lists it on the Quadro RTX 6000 and RTX A6000, but not on the RTX 6000 Ada Generation or RTX PRO 6000 Blackwell. CGChannel reported NVIDIA's memory-pooling claim for bridged Quadro RTX 6000 cards, and PNY's A6000 bridge listing at B&H also describes pooling. Treat these as vendor claims and confirm support for your application before planning around combined memory. CoreWeave's own documentation for its RTX PRO 6000 Blackwell instance says plainly that the card "lacks NVLink for memory pooling across GPUs." A newer card in this family is not a more connected one: the site spec table's FP32 ratings of 91.1 TFLOPS for RTX 6000 Ada Generation and 38.7 TFLOPS for RTX A6000 give a derived ratio of about 2.4 times. These are vendor claims for peak throughput, not measured application speed; the card with the higher rating lacks the older card's NVLink option.
How to tell which one you are being offered
- For the four RTX families above, check memory size first. 24GB is the Quadro RTX 6000. 48GB is either the RTX A6000 or the RTX 6000 Ada Generation, so you need a second clue. 96GB is the RTX PRO 6000 Blackwell, and nothing else in this family comes close, which answers the "rtx 6000 vram" question on its own for that card.
- Ask for the architecture name, not just the model number. Turing means the Quadro RTX 6000, Ampere means the RTX A6000, Ada Lovelace means the RTX 6000 Ada Generation, and Blackwell means the RTX PRO 6000. Ask the seller to confirm the full model against the spec table.
- Check whether NVLink is present. Only the Quadro RTX 6000 and the RTX A6000 have it. A claim of NVLink bridging on either newer card conflicts with the spec table.
- Do not trust the listing title alone. Retail titles mix these names, so read the part number and the memory size together: a "Quadro RTX 6000" listed with 48GB cannot be the 2018 card, which has 24GB, and 48GB alone still cannot separate the A6000 from the Ada card.
- Use dated purchase prices as a sanity check, not proof. Aggregator GPUDojo showed a used Quadro RTX 6000 from $750 and a used RTX A6000 from $4,400, both as of 27 September 2026, and Newegg listed an RTX 6000 Ada Generation card at $9,599.00 and an RTX PRO 6000 Blackwell Workstation Edition card at $19,999.99, both on 28 September 2026. These are asking prices on the dates shown. Price narrows the field but cannot prove the model.
What each one costs to rent right now
The live table below shows the cheapest current on-demand price for each card, its provider, and how many providers have it in stock. If you searched "rtx 6000 price" expecting a single number, check the full model name before comparing offers.
| GPU | Cheapest $/GPU-hr | Provider | Providers in stock |
|---|---|---|---|
| Quadro RTX 6000 | none in stock | ||
| RTX A6000 | $0.44 | LeaderGPU | 5 |
| RTX 6000 Ada Generation | $0.78 | QuantaCloud | 4 |
| RTX PRO 6000 Blackwell | $0.59 | RunPod | 5 |
Which one to rent or buy
Match the card to the job.
For a single-GPU workload needing more than 48GB and fitting within 96GB, choose the RTX PRO 6000 Blackwell. It is the only one of the four with native FP4, although peak figures do not predict your application's speed. NVIDIA's inference guide explains that context length and batch size increase KV-cache memory; Hugging Face's memory guide also accounts for training gradients, optimizer states and activations. Those requirements make a single parameter-count cutoff unreliable. According to ai-infrastructure.net, RTX PRO 6000 uses NVIDIA's RTX Enterprise driver branch and carries ISV certifications for Maya, SolidWorks and Creo that GeForce cards do not carry. Check the certification for your specific application before relying on it for rendering as well as AI. Price is the argument for renting: Newegg asked $19,999.99 for the Workstation Edition on 28 September 2026, and the tracker videocardprices.com recorded the same SKU anywhere from $12,379.99 on 13 July 2026 to $19,999.00 on 12 September 2026, too unstable to plan a purchase around. NVIDIA sells Workstation, Max-Q Workstation and Server Editions, the last passively cooled; CDW listed the Server Edition card at $13,824.99 on 21 September 2026. The RTX PRO 6000 buying guide covers the editions in full.
For a workload that fits in 48GB, compare the RTX A6000 and RTX 6000 Ada Generation. Between the two, start with Ada for its higher specified FP32 throughput, and consider the RTX A6000 if you specifically need NVLink. PNY's bridge listing describes combined memory, but confirm that your application supports the intended setup before choosing it. Compare the families on the RTX 6000 Ada versus RTX A6000 page and the RTX A6000 versus RTX PRO 6000 Blackwell page.
Buying used makes sense only at the bottom of the range, for a legacy workload you already know the card runs. Tom's Hardware reported a $6,300 launch price for the Quadro RTX 6000 in 2018, and GPUDojo showed used cards from $750 on 27 September 2026. Rent one by the hour to confirm your workload first.
Check the rent versus buy calculator against your own hours-per-month figure before committing either way, and run your model size through the LLM VRAM calculator before choosing a card. The decision rule for a single GPU: if the full workload needs more than 48GB but fits within 96GB, choose the RTX PRO 6000 Blackwell. If it fits in 48GB, start with the RTX 6000 Ada Generation for its higher specified FP32 throughput versus A6000; choose the RTX A6000 when you need its NVLink option. If it exceeds 96GB, reject all four as single-card solutions.
Sources
- NVIDIA newsroom: new Quadro graphics solutions bring Fermi to market, 27 July 2010
- CGChannel: NVIDIA unveils the Quadro M6000, March 2015
- TechPowerUp: NVIDIA unveils the Quadro M6000 refresh, retrieved 28 September 2026
- NVIDIA Quadro SLI HB / NVLink bridge quick start guide, retrieved 28 September 2026
- NVIDIA newsroom: NVIDIA reinvents computer graphics with Turing architecture, 13 August 2018
- Tom's Hardware: NVIDIA Quadro RTX 6000 price and specs, retrieved 28 September 2026
- CGChannel: NVIDIA unveils new RTX A6000 and A40 Pro Ampere GPUs, October 2020
- VideoCardz: NVIDIA announces RTX A6000 workstation graphics card, now available, retrieved 28 September 2026
- ServeTheHome: NVIDIA RTX A6000 48GB GPU drops the Quadro designation, retrieved 28 September 2026
- Notebookcheck: NVIDIA RTX 6000 Ada Generation announced ahead of December release, retrieved 28 September 2026
- VideoCardz: NVIDIA RTX A6000 "Ada" workstation GPU officially costs $6,800, retrieved 28 September 2026
- NVIDIA newsroom: NVIDIA Blackwell RTX PRO comes to workstations and servers, 18 March 2025
- CoreWeave docs: RTX PRO 6000 Blackwell Server Edition (High Memory) instance, retrieved 28 September 2026
- Nebius: RTX PRO 6000 Blackwell Server Edition compute page, retrieved 28 September 2026
- Google Cloud blog: Cloud Run supports NVIDIA RTX 6000 PRO GPUs, 2 February 2026
- NVIDIA TensorRT support matrix, updated 8 September 2026
- SemiAnalysis: NVIDIA Tensor Core evolution from Volta to Blackwell, 23 June 2025
- GPUDojo: Quadro RTX 6000 prices, as of 27 September 2026
- GPUDojo: RTX A6000 prices, as of 27 September 2026
- Newegg: PNY RTX 6000 Ada 48GB graphics card listing, retrieved 28 September 2026
- Newegg: NVIDIA Blackwell RTX PRO 6000 96GB graphic card listing, retrieved 28 September 2026
- videocardprices.com: NVIDIA RTX PRO 6000 Blackwell price tracker, as of 27 September 2026
- CDW: PNY NVIDIA RTX PRO 6000 graphic card, 96 GB GDDR7, Server Edition listing, retrieved 21 September 2026
- ai-infrastructure.net: RTX workstation cards versus GeForce for AI, retrieved 28 September 2026
- Puget Systems: NVLink bridge compatibility chart, retrieved 28 September 2026
- NVIDIA: RTX PRO 6000 family editions, retrieved 28 September 2026
- NVIDIA: RTX PRO 6000 Workstation Edition, retrieved 28 September 2026
- NVIDIA: RTX PRO 6000 Server Edition, retrieved 28 September 2026
- CGChannel: NVIDIA unveils Quadro RTX cards, August 2018
- B&H: PNY RTX A6000 NVLink bridge listing, retrieved 28 September 2026
- NVIDIA: inference memory and KV cache, 17 November 2023
- Hugging Face: model memory anatomy, accessed 21 September 2026