The RTX PRO 6000 puts 96 GB of VRAM on a single Blackwell card, and that is the reason to look for RTX PRO 6000 cloud capacity. Models that would otherwise need two or more GPUs, with all the complexity of splitting them, fit on one. This article explains the three PRO 6000 variants you will see on Cloudj, what the memory makes possible, how to search for one by country, and the software check you need before relying on a Blackwell GPU.
For the general method of choosing GPUs by country, see the GPU cloud by location guide.
Three variants on the list
Cloudj lists three RTX PRO 6000 models. All share the Blackwell architecture and 96 GB of VRAM; they differ mainly in form factor and power budget, which you can see on each model page.
| Variant | Model page | What it is |
|---|---|---|
| RTX PRO 6000 Workstation | rtx-pro-6000-ws | Full-power workstation card |
| RTX PRO 6000 Max-Q | rtx-pro-6000-max-q | Lower-power version of the workstation card, often used in dense multi-GPU builds |
| RTX PRO 6000 Server | rtx-pro-6000-s | Passively cooled server edition for data-centre chassis |
For most rental workloads, the differences matter less than the 96 GB they share. Sustained performance can differ with the power budget, so for long jobs compare a short benchmark on each if more than one is listed. In practice, search all three and choose by price, reliability and max duration.
These are workstation-class GPUs. On Cloudj they need a higher account level than new accounts start with; if the console says the class is not available at your level, contact support from the console.
What 96 GB on one card enables
As general guidance (exact needs depend on precision, context length and engine):
- 70B-class LLMs in 4-bit or 8-bit on a single GPU, with room for a useful context window, instead of splitting across cards.
- Long-context inference for mid-size models, where the KV cache rather than the weights is what fills memory.
- Large image and video generation models at full or near-full precision without offloading tricks.
- Fine-tuning of mid-size models with larger batches, or LoRA on larger base models.
- Big rendering scenes whose textures and geometry exceed what 24 or 48 GB cards can hold.
A single large card is simpler than several smaller ones: no tensor parallelism, no inter-GPU communication, fewer failure points. If your model needs even more memory than 96 GB, the 141 GB class H200 NVL is the next step up; the H200 cloud rental article covers it. If you only need around 32 GB, the RTX 5090 in the USA article covers a smaller Blackwell option.
Searching by region
GPU location on Cloudj is a country filter over a live list of host machines. Open the RTX PRO 6000 deploy page and enter the countries you accept in the Country field. A few common starting points:
| Where you work | Example codes |
|---|---|
| North America | US |
| Europe | DE, NL |
| East Asia | JP |
| Batch work, anywhere | No country filter |
Then switch the model between the Workstation, Max-Q and Server variants, or better, clear the model and set a minimum VRAM of 96 GB to see all of them at once with your country codes.
Each result shows the country (never a city), GPU count, vCPU, RAM, download speed, reliability, max duration, bandwidth price and a Datacenter badge where relevant. If a machine is listed, you can deploy it. Stock changes during the day; large workstation cards can be scarce in any single country, so be ready to widen your codes or remove the filter for jobs that are not latency-sensitive. Live "from" prices are on each model page.
Blackwell driver and software check
Blackwell needs recent drivers and recent builds of CUDA libraries. Older PyTorch versions and some inference or attention libraries may not include kernels for it. Run these checks right after you connect:
nvidia-smi
python -c "import torch; print(torch.__version__, torch.version.cuda); print(torch.cuda.get_device_name(0)); print(torch.randn(2, 2, device='cuda') @ torch.randn(2, 2, device='cuda'))"If PyTorch reports that no kernel image is available, install a build for a recent CUDA version. Check the release notes of every library in your stack for Blackwell support. The NVIDIA CUDA documentation lists toolkit versions and supported architectures.
The cleanest path is a custom image built on a recent CUDA base, so every deploy starts from a known-good environment. Custom images must be public Docker images and, like workstation GPUs, need a higher account level. The custom Docker image for GPU article walks through building one.
Operating notes
- Disk: a 96 GB card invites large models. Size the disk for weights, caches and outputs; it cannot be resized later.
- Downloads: pull model weights on the host from their source rather than uploading them.
- End date: each machine shows "available until" a date, after which the instance and disk are removed. Plan backups well before.
- No SLA: machines are independent hosts. Keep your setup scripted so a new machine can replace the old one quickly.
- Licences: large open models often come with specific licence terms; check them for your use.
FAQ
Which RTX PRO 6000 variant should I pick?
All three have 96 GB. Search them together by setting a minimum VRAM of 96 GB, then choose on price, reliability and end date.
Can a 70B model run on one RTX PRO 6000?
In 4-bit or 8-bit, generally yes, with room for context. Exact headroom depends on context length and engine.
Why can I not select an RTX PRO 6000?
Workstation-class GPUs need a higher account level. The console tells you, and you can contact support from there.
Will my current CUDA image work on it?
Only if it is recent enough for Blackwell. Run the checks above before starting a long job.
Find a 96 GB card
Open the RTX PRO 6000 deploy page, set your countries or a 96 GB minimum, and see what is listed now. Compare all live prices on GPU cloud.