The RTX 5090 is the first consumer card with 32 GB of VRAM, and that extra 8 GB over the RTX 4090 is the main reason people look for RTX 5090 cloud USA capacity. It lets a single card hold models that were just out of reach on 24 GB. This article covers what the extra memory buys you, how to search for a 5090 in the United States on Cloudj, and the one technical check you must not skip: making sure your software supports the Blackwell architecture.
For US searches in general, including other models and VPS pairing, see the GPU cloud USA guide.
What 32 GB adds over 24 GB
On paper, the RTX 5090 model page lists 32 GB of VRAM, the Blackwell architecture and higher memory bandwidth than the RTX 4090. In practice, the memory is what changes which jobs fit. As general guidance:
| Workload | RTX 4090 (24 GB) | RTX 5090 (32 GB) |
|---|---|---|
| 13B LLM in 8-bit with long context | Tight | Comfortable |
| 30B to 34B LLM in 4-bit | Fits with short context | Fits with more room for context and batching |
| Large image models at full precision | Often needs offloading or lower precision | More often fits as is |
| Open video generation models | Many need aggressive memory tricks | Several become practical on one card |
| LoRA fine-tuning with larger batches | Limited | More room for batch size or sequence length |
These are rough expectations; your exact needs depend on context length, batch size, precision and the inference engine. The rule of thumb: if your job runs on a 4090 with room to spare, a 5090 mainly adds speed. If your job barely fits or does not fit on a 4090, the 5090's memory is the real upgrade. The RTX 4090 vs RTX 5090 cloud article covers this trade-off in detail.
Searching for an RTX 5090 in the US
GPU location on Cloudj is a country filter over a live list of host machines. Open the RTX 5090 deploy page, which preselects the model, and set the Country field to US.
Each result shows the country (never a state or city), GPU count, VRAM, vCPU, RAM, download speed, reliability, max duration and bandwidth price, plus a Datacenter badge where relevant. The list is sorted by total price.
If a machine is listed, you can deploy it. If not, stock changes during the day; check again later or widen the search:
- Add Canada:
US, CA. The North America GPU guide covers both countries. - Clear the model and set a minimum VRAM of 32 GB to see other cards that would fit, including 48 GB class workstation cards if your account level allows them.
- For batch jobs with no location need, remove the country filter entirely.
Note that a US result could be anywhere in the country. If you need a specific coast for latency, test the machine after deploying rather than assuming. Live "from" prices for the card are on the model page.
The Blackwell compatibility check
This is the step that catches people out. Blackwell is a newer architecture than Ada (RTX 4090) and needs recent software. Older CUDA toolkits and older PyTorch builds may not include kernels for it, which shows up as errors such as "no kernel image is available for execution on the device" or a framework that silently falls back to CPU.
Before committing to a long job:
- Check the driver and CUDA version on the host:
nvidia-smiThe header shows the driver version and the highest CUDA version it supports.
- Check your framework sees the card and can run on it:
python -c "import torch; print(torch.__version__, torch.version.cuda, torch.cuda.get_device_name(0)); print(torch.ones(1, device='cuda') * 2)"If this fails, install a PyTorch build made for a recent CUDA version. The PyTorch installation selector shows the current options.
- Check your other libraries. Inference engines, attention kernels and quantization libraries each have their own CUDA builds. Read their release notes for Blackwell support before you rely on them.
If you use a custom Docker image, build it on a recent CUDA base so it works on Blackwell from the start. The custom Docker image for GPU article explains how to prepare one. Custom images must be public and need a higher account level; the default Ubuntu + CUDA and PyTorch + JupyterLab templates are available to everyone.
Running it well
- Access: SSH key only. The Connection tab shows the host and port:
ssh -p PORT root@HOST. - Disk: choose generously at deploy; it cannot be resized later, and large models fill it fast.
- Availability end: each machine shows "available until" a date. Keep checkpoints and outputs off the instance well before then.
- Stop vs destroy: stopping keeps the disk but releases the GPU; starting again waits until it is free. Destroy when you are done.
- Billing: per minute from your prepaid balance, plus bandwidth per GB. The deploy price is locked for the instance.
When a 5090 is not the right choice
If your model needs more than 32 GB at the precision you want, a single 5090 will not do. Look at 48 GB class cards, the 96 GB RTX PRO 6000 or datacenter cards such as the H100, which need a higher account level. If your workload already runs comfortably on 24 GB and you care about cost per job, compare live prices against a 4090 before choosing.
FAQ
Is the RTX 5090 always available in the US?
No. Stock changes during the day. Select the 5090, set the Country to US, and deploy if a machine is listed.
Will my existing PyTorch setup work on a 5090?
Only if it is recent enough to support Blackwell. Run the check above before starting a long job.
Can I choose a US state or city?
No. GPU location is country level only.
Can I use more than one RTX 5090?
The GPU count selector offers 1×, 2×, 4× and 8× in one machine. Multi-GPU listings appear only when a host has them.
Find a US RTX 5090
Open the RTX 5090 deploy page, set the Country to US and see what is listed now. Compare all live prices on GPU cloud.