Cloud GPU rendering: rent RTX and L40S cards with crypto for Blender and GPU renderers
For ray-traced rendering the most expensive data-center card is often the wrong one, and the right one costs under a dollar an hour here, paid in crypto only. Here is which card fits which scene, how to render from the command line, what an animation costs, and when a managed render farm is the better deal.
The best cloud GPU for rendering is a card with RT cores and enough memory for your scene: an RTX 4090 (24 GB, $0.34 an hour) for most Blender Cycles, Redshift and Octane work, an RTX 5090 (32 GB, $0.49) for heavier scenes, and an L40S (48 GB, $0.87) when a scene does not fit in 32 GB. The A100, H100, H200 and B200 are built for AI and have no RT cores, so they are a poor and expensive choice for ray tracing. On a rented machine you render over SSH from the command line, and several cards can render different frames at the same time.
Best GPU for 3D rendering, by scene size
Each card must hold the whole scene: geometry, textures and the render buffers. GeForce and L40S cards do not pool their memory, so a machine with four cards still has the memory of one per frame.
| Card | Memory | RT cores | Per hour | Good for |
|---|---|---|---|---|
| RTX 4090 | 24 GB | 3rd generation | $0.34 | Most product, archviz and motion scenes, at the lowest hourly price of the three |
| RTX 5090 | 32 GB | 4th generation | $0.49 | Heavier scenes and textures; 1,792 GB/s of memory bandwidth |
| L40S | 48 GB | 3rd generation | $0.87 | Large scenes, 8K textures, many instanced assets |
| A100, H100, H200, B200 | 80 to 180 GB | None | $0.99 to $4.89 | AI training and inference, not ray tracing |
RT core generations from NVIDIA's product pages. Without RT cores, renderers fall back to slower ray tracing on the general CUDA cores.
Renderers on a rented Linux machine
Our machines run Ubuntu and are reached over SSH, without a desktop. What matters is whether your renderer has a Linux command-line version, and how its licence works on a remote machine.
| Renderer | Runs headless on Linux | Licence |
|---|---|---|
| Blender Cycles (OptiX or CUDA) | Yes, from the command line | Free and open source |
| Redshift, Octane, V-Ray GPU | Depends on the host application and the vendor's Linux build | Commercial: check that the vendor allows use on a cloud machine and how you sign in |
| Unreal Engine, Lumion, D5, Twinmotion | Built around a desktop and Windows | Better served by a render farm or a Windows workstation |
Applications such as 3ds Max and Cinema 4D are desktop programs; on a rented Linux machine you can only use renderers that ship a command-line version for Linux.
Render a Blender animation from the command line
Download the Linux build of Blender from blender.org onto the machine, upload your .blend file with its textures packed, and render with OptiX. On a machine with several cards, start one Blender process per card, each on its own range of frames: that uses every card fully, without the scene having to fit across them.
# From your computer: upload the scene (pack textures first: File > External Data > Pack Resources)
rsync -avP scene.blend root@<address>:~/render/
# On the machine: render frames 1 to 250 with OptiX on one card
cd ~/render
~/blender/blender -b scene.blend -E CYCLES -o //frames/frame_#### -s 1 -e 250 -a -- --cycles-device OPTIX
# On a 2-card machine: one process per card, half the frames each
CUDA_VISIBLE_DEVICES=0 ~/blender/blender -b scene.blend -E CYCLES -o //frames/frame_#### -s 1 -e 125 -a -- --cycles-device OPTIX &
CUDA_VISIBLE_DEVICES=1 ~/blender/blender -b scene.blend -E CYCLES -o //frames/frame_#### -s 126 -e 250 -a -- --cycles-device OPTIX &
wait
# From your computer: download the frames
rsync -avP root@<address>:~/render/frames/ ./frames/What an animation costs
A 250-frame animation (10 seconds at 25 frames per second), by the minutes each frame takes on one card. Render a test frame, then read across.
| Card, per hour | 1 min per frame | 3 min per frame | 10 min per frame |
|---|---|---|---|
| RTX 4090, $0.34 | $1.42 | $4.25 | $14.17 |
| RTX 5090, $0.49 | $2.04 | $6.13 | $20.42 |
| L40S, $0.87 | $3.63 | $10.88 | $36.25 |
Cost = frames × minutes per frame ÷ 60 × hourly price. Eight cards render the same animation about eight times sooner for the same total cost, since each frame is rendered once. Every price is on what a GPU-hour costs.
Render hours are bought in blocks and paid in crypto only: 75 RTX 4090 hours for $25.50, 60 RTX 5090 hours for $29.40 or 30 L40S hours for $26.10. Unused hours stay on your account for the next project. See the coins and networks we accept.
Rented GPU or render farm
Managed render farms and rented machines solve different problems.
| Managed render farm | Rented GPU machine | |
|---|---|---|
| How you work | Upload the scene through a plugin or website; the farm renders and returns frames | SSH into your own machine, install what you need, run the render yourself |
| Software | Windows applications and plugins already installed, licences often included | Linux command-line renderers; you bring commercial licences |
| Pricing | Per node-hour, GHz-hour or benchmark unit, set by each farm | Per GPU-hour, billed by the minute |
| Best for | Cinema 4D, 3ds Max, Unreal and deadline-driven jobs with many plugins | Blender, repeated renders of the same pipeline, and mixing rendering with AI work |
When renting by the hour pays off
A rented machine makes most sense when you render often with the same setup: install Blender and your add-ons once, keep them on the machine's disk, and stop the machine between jobs so it costs nothing while idle. It also suits studios that use the same cards for AI image work during the day and rendering at night. For a one-off job in Cinema 4D or 3ds Max with a dozen plugins, a managed farm saves the setup. Whichever you choose, rent GPUs by the hour only for the minutes you render.
GPU rendering questions
What GPU is best for rendering?
A card with RT cores and enough memory for the scene. In the cloud that means an RTX 4090 or RTX 5090 for most work, and an L40S with 48 GB for large scenes. Data-center AI cards without RT cores render slower per dollar.
Is rendering CPU or GPU heavy?
It depends on the renderer. Cycles, Redshift, Octane and V-Ray GPU run mostly on the GPU and are much faster there; CPU renderers such as Arnold on CPU or Corona use the processor instead.
Are render farms still a thing?
Yes. Managed farms remain the easiest way to render Windows applications with many plugins on a deadline. Rented GPU machines have become the cheaper option for Blender and for teams comfortable with the command line.
How do I enable GPU rendering in Blender?
In Preferences > System, choose OptiX and tick your GPU, then set Device to GPU Compute in the render properties. On a headless machine, add -- --cycles-device OPTIX to the command line instead.
Sources
- Blender manual, command line renderingread 19 November 2026
- Blender manual, GPU renderingread 19 November 2026
- NVIDIA L40Sread 19 November 2026
- NVIDIA GeForce RTX 5090read 19 November 2026