The RTX 4090 vs 5090 cloud question is not "which card is faster". The 5090 is the newer card and generally quicker, but when you rent by the minute, what matters is whether your job fits in memory, whether your software runs on the architecture, and what each finished job costs at today's prices. This article walks through those three questions and then covers how to find each card in the region you need.
For the broader choice between consumer and datacenter cards, see consumer vs datacenter GPU cloud.
RTX 4090 vs 5090 cloud specs side by side
| RTX 4090 | RTX 5090 | |
|---|---|---|
| Architecture | Ada Lovelace | Blackwell |
| Memory | 24 GB | 32 GB |
| Class | Consumer | Consumer |
| Software maturity | Supported by essentially every CUDA framework release of recent years | Needs recent CUDA and framework builds |
| Account level | Available to new accounts | Available to new accounts |
There is also an RTX 5090 D on the GPU list, a regional variant with 32 GB. If it is listed where you need it, treat it as a 5090-class option and check its model page for details. Specifications beyond these are on each model page; we do not quote benchmark results here, because your model and settings decide the real difference.
Question 1: does the job fit in 24 GB?
The 8 GB difference matters more than the speed difference for some jobs.
- Fits comfortably in 24 GB: SDXL generation, most LoRA training on 7B-8B models, quantized 13B-class inference, vision models, rendering most scenes. Either card works; the choice is about cost and speed.
- Tight on 24 GB: quantized models in the 30B range with useful context, larger image or video generation models, LoRA with long sequences. The 5090's 32 GB can remove offloading or tiling, which can matter more than raw speed.
- Does not fit on either: larger models at higher precision. Look at 48 GB or larger cards instead.
If you hit "CUDA out of memory" on a 4090 and are fighting it with workarounds, try the 5090 before anything else.
Question 2: does your software run on Blackwell?
New architectures need new software. For the 5090:
- use a recent CUDA toolkit and driver that support Blackwell;
- use a PyTorch build compiled for it (recent releases are; older pinned versions may not be). The PyTorch install page shows current builds;
- check custom CUDA extensions, attention kernels and quantization libraries, which sometimes lag behind;
- watch for errors such as "no kernel image is available for execution on the device", which usually mean a library was not built for the card.
A quick check after deploy:
nvidia-smi
python -c "import torch; print(torch.__version__, torch.cuda.get_device_name(0), torch.cuda.get_device_capability(0))"If you deploy a custom image, rebuild it on a current CUDA base before switching to the 5090. Custom images need a higher account level. The 4090 has no such concerns: almost any image from recent years runs.
Question 3: what does each finished job cost?
Rent by the minute and think in cost per job, not price per hour:
cost per job = locked price per minute x minutes to finish
+ storage + bandwidthThe 5090 usually has a higher hourly price and a shorter run. Which is cheaper depends on how much faster your job actually runs. The reliable way to know:
- Deploy each card for a short test with your real workload.
- Measure throughput: steps per minute, images per minute or tokens per second on your own settings.
- Multiply out to the full job with each card's live price.
- Destroy both test instances.
Per-minute billing keeps this test cheap. Live "from" prices are on each model page, and the console shows the exact price before you deploy; that price is then locked for the instance. The GPU cost estimation guide covers storage and bandwidth in the same calculation.
Finding each card where you need it
Cloudj GPUs come from a live inventory of independent host machines, and which models are listed in which country changes during the day. Never assume one card is easier to find than the other in a given place; search.
Open the RTX 4090 deploy page, enter the country codes you need, note the results, then switch the model to the 5090 and search again. If one card is not listed:
- add neighbouring country codes;
- loosen reliability or price filters;
- for batch jobs, clear the Country field;
- check back later.
Two regional articles go further: RTX 4090 in Japan and RTX 5090 in the USA.
A simple decision rule
- Pick the 4090 when the job fits in 24 GB, your software stack is older or pinned, or your test shows little speed gain from the 5090.
- Pick the 5090 when 24 GB forces offloading or tiling, your stack supports Blackwell, or your test shows a cost-per-job advantage.
- Pick whichever is listed when the job is urgent and both would work.
Whichever you choose, the instance runs on an independent host with no hardware SLA, has an "available until" date after which instance and disk are removed, and should be destroyed when the job is done.
FAQ
Is the RTX 5090 always better than the 4090?
It is newer and has more memory, but the cheaper option per job depends on your workload and live prices. Test both.
Will my existing Docker image work on a 5090?
Only if its CUDA and framework versions support Blackwell. Older images may need rebuilding.
What is the RTX 5090 D?
A regional variant of the 5090 with 32 GB, listed separately on the GPU list.
Can a new account rent both?
Yes. Both are consumer GPUs, available from the lowest account level.
Test both cards
Open the RTX 4090 deploy page, run a short test, then repeat on the 5090. Compare live prices for both on GPU cloud.