RTX 5090
from $0.36/hr.
32 GB of GDDR7 at 1.79 TB/s and fifth-generation Tensor Cores with FP4. Blackwell for fast inference on small models and image generation.
BlackwellAvailable now
32GB GDDR7
Billed from your wallet · stop any time
Memory32GB
Bandwidth1,792GB/s
ArchitectureBlackwell
Form factorPCIe 5.0
Good forWhat the RTX 5090
What the RTX 5090
is built for.
Tensor-core precisions: FP4, FP8, BF16, FP16 and TF32.
Best at
- Serving 7B–13B models
- LoRA and QLoRA on models up to 13B
- Flux and SDXL image generation
- FP4 and FP8 inference
Launch from a terminal
Same card, same price as the console.
$ packet launch --gpu rtx5090 --wait
CLI docs SizingWill my model fit
Will my model fit
on 32 GB?
Rules of thumb for weights plus a modest KV cache or optimizer state. Longer contexts and bigger batches need more.
| Workload | Needs | On a RTX 5090 |
|---|---|---|
| Llama 3.1 8B · FP16 inference | ~18 GB | Fits |
| Qwen 32B · 4-bit inference | ~20 GB | Fits |
| Qwen 32B · FP8 inference | ~36 GB | Too big |
| Llama 70B · 4-bit inference | ~42 GB | Too big |
| Llama 8B · LoRA fine-tune (BF16) | ~24 GB | Fits |
| 13B · QLoRA fine-tune | ~16 GB | Fits |
| Flux.1 [dev] · BF16 | ~33 GB | Tight |
| SDXL · FP16 | ~12 GB | Fits |
Other GPUsNot the right fit?
Not the right fit?
Compare the fleet.
These are on sale now, billed hourly with no contract.
Launch a RTX 5090
from $0.36/hr.
Sign up, top up, launch. You'll be on a GPU in minutes.
Launch a RTX 5090No contractBilled hourlyTalk to a human any time
