How Much VRAM Do You Need for 3D Rendering? (2026 Guide)
There is no single number: the VRAM you need is the peak memory of your heaviest frame, plus some headroom. Measure it, then compare it with the memory of one card, because GPU memory is not pooled.
- Measure: render the heaviest frame and read Peak memory in Blender, or watch Memory-Usage in nvidia-smi
- Compare with one card: 8 GB (RTX 4060), 12 GB (RTX 4070/5070), 16 GB (RTX 5080), 24 GB (RTX 4090), 32 GB (RTX 5090)
- More GPUs: add speed, not memory: every GPU keeps its own copy of the scene
- Too big: optimize textures and instances, or let Redshift, Octane or Cycles use system RAM (out-of-core), more slowly
“How much VRAM do I need?” is the wrong first question; the right one is “how much does my scene use?” Two artists with the same software can need 6 GB or 30 GB depending on textures, geometry and effects. This guide shows how to measure your scene, how to compare it with the cards on the market, and what happens when it does not fit.
Checked against NVIDIA, Maxon, OTOY and Blender documentation, November 2026.
1. What uses VRAM when you render
- Textures. Memory is width × height × channels × bytes per channel: an 8-bit RGBA 4K image is 64 MiB, the same image as 32-bit float 256 MiB, an 8K float image 1 GiB.
- Geometry and its acceleration structure (BVH): every triangle, plus subdivision and displacement created at render time. Instances share one mesh.
- Volumes, hair and particles, which can be large on their own.
- Frame buffers for the image and every render pass (AOV) at the output resolution.
The total at the busiest moment of the heaviest frame is your peak. That is the number to plan around.
2. How to measure your scene
- Blender: render the heaviest frame with F12. The render window header shows Mem and Peak; Peak is the number you need.
- Any engine: while it renders, run nvidia-smi in a Command Prompt and read Memory-Usage for each GPU. Redshift and Octane also report memory in their own log or statistics.
- Read one card. On a multi-GPU machine every GPU shows about the same usage, because each holds its own copy. Compare that figure with the memory of one card.
- Leave headroom. If the peak is within about 10% of the card, a bigger frame or the next shot can push it over.
How to read the figure. The top line is what Blender prints after a render: badge 1 is the Peak value, the most memory the frame needed at any point. The black panel is nvidia-smi during a render on four GPUs: badge 2 shows that each card is using about the same 14.8 GB, and badge 3 is the reason, each card holds the whole scene. So this scene needs a card with more than about 15 GB, whether you have one GPU or eight. The numbers here are an example, not a measurement.
3. Which cards fit which scenes
Memory per card, from NVIDIA’s specifications: GeForce RTX 4060 8 GB, RTX 4070 and RTX 5070 12 GB, RTX 5080 16 GB, RTX 4090 24 GB, RTX 5090 32 GB. Because memory is not pooled, four RTX 4090s still give you 24 GB per scene; they render faster, but they do not hold a bigger scene. Maxon states this for Redshift, and the Blender Manual says the same for Cycles: each GPU can only use its own memory.
How to read the figure. Each row is an example scene by its peak memory, each column a card. Green means the scene uses less than 90% of one card, amber that it fits with little headroom, red that it does not fit. The last column is four RTX 4090s: it matches the single RTX 4090 column, because the memory of one card is still the limit. Only the 32 GB RTX 5090 holds the 28 GB scene; the 40 GB scene needs optimizing or out-of-core rendering.
4. When the scene does not fit: out-of-core
Several GPU renderers can keep going when VRAM runs out by placing part of the data in system RAM. It works, but data in system RAM is slower to reach than data on the card.
- Redshift moves geometry and textures that do not fit out of core, controlled in its memory settings.
- Octane has out-of-core for textures and geometry in its Out of Core settings; that system memory is shared by the GPUs.
- Cycles (CUDA and OptiX) moves data to system RAM when VRAM is full and leaves at least half of the RAM for other programs. Textures moved out cost a moderate slowdown; other scene data much more.
Out-of-core is a safety net, not a plan: if a scene always runs out of core, a card with more memory or a lighter scene renders it faster.
5. Optimize before you buy
The cheapest VRAM is the VRAM you do not use. Size textures to how big they appear on screen and keep 8-bit images for color, instance repeated objects instead of copying them, and only render the passes you need. Our guide to optimizing Blender scenes shows the settings, and the same ideas apply to Redshift and Octane.
VRAM planning checklist
Tick each item before you choose a card or a server.
SaaS or IaaS render farm for a scene near the VRAM limit?
A scene that fits in 24 GB on one machine fits in 24 GB on any machine; what changes between farms is which cards they offer and whether you can measure on them before you commit.
VRAM limits and render farms
How the two farm models fit a scene close to the memory of one card.
| Your situation | SaaS render farm | IaaS render farm (iRender) |
|---|---|---|
| You know the scene’s peak and it fits the farm’s cards | A good fit: upload and the farm renders the range | Works as well: open the scene and render |
| You do not know whether it fits in 24 or 32 GB | Check the GPU memory the farm lists before you submit | Open the scene on an RTX 4090 (24 GB) or RTX 5090 (32 GB) server and read the peak yourself |
| The scene needs out-of-core | Check whether the farm enables it and how much RAM its nodes have | Every server has 256 GB of ECC RAM, and you set out-of-core yourself |
A VRAM test on an iRender server
How to read the figure. The left box is your own workstation and the right box is the iRender server: the same desktop, software version, plugins, licenses and scene settings. The only difference is the GPU, 24 GB per card on RTX 4090 servers and 32 GB per card on RTX 5090 servers, so you can measure the peak of your own scene on the card you are considering.
- Upload first. Put the project in iRender Drive before you start a server; transfer and storage are free and no rental time runs while you upload.
- Pick the card. A 1-GPU server (3S with RTX 4090 or 3i with RTX 5090) is enough to measure memory; 1-GPU servers have a one-hour minimum.
- Open the scene. Cinema 4D, Redshift and Octane are pre-installed; install Blender, Houdini or Maya yourself. They stay in your server image.
- Measure the heaviest frame with the steps in section 2.
- Render. If it fits, render on a 2x, 4x or 8x server for speed; multi-GPU servers bill per minute.
Frequently asked questions
1. How much VRAM do I need for 3D rendering?
As much as the peak memory of your heaviest frame plus about 10% headroom. Render that frame and read Peak in Blender or Memory-Usage in nvidia-smi, then compare it with the memory of one card.
2. Do multiple GPUs add up their VRAM?
No. In Redshift, Octane and Cycles each GPU keeps its own copy of the scene, so four 24 GB cards still hold a 24 GB scene. More GPUs make rendering faster.
3. Is 8 GB of VRAM enough for rendering?
For light scenes, yes. Measure your own scenes: textures, displacement, volumes and many render passes quickly push the peak past 8 GB.
4. How much VRAM does a 4K texture use?
An 8-bit RGBA 4K texture uses 64 MiB; as 32-bit float it uses 256 MiB.
5. What happens when a scene runs out of VRAM?
Redshift, Octane and Cycles can move part of the data to system RAM (out-of-core) and keep rendering, more slowly. Other renderers may stop with an out-of-memory error.
6. How much VRAM do the RTX 4090 and RTX 5090 have?
The RTX 4090 has 24 GB of GDDR6X and the RTX 5090 has 32 GB of GDDR7.
7. How do I check VRAM usage while rendering?
Run nvidia-smi in a Command Prompt during the render and read Memory-Usage for each GPU. In Blender, the render window also shows Mem and Peak.
Render heavy scenes on iRender
iRender is an IaaS GPU render farm: you rent a whole bare-metal server with 1, 2, 4 or 8 NVIDIA RTX 4090 (24 GB) or RTX 5090 (32 GB) GPUs, an AMD Ryzen™ Threadripper™ PRO CPU, 256 GB ECC RAM and a 2 TB NVMe SSD, connect by remote desktop and work on it like your own workstation, with full administrator access. Cinema 4D, Redshift and OctaneRender are pre-installed; your server image is saved between sessions, and the physical machine is fully wiped before anyone else uses it.
Multi-GPU servers are billed per minute (1-GPU servers have a one-hour minimum), and sessions of 3 hours or more get 10% off. Our 4-GPU and 8-GPU servers are liquid-cooled; in our own measurements the GPUs hold 60–70°C under sustained render load. Elite support is available 7 days a week by hotline, live chat, Skype, WhatsApp and email.
New users get a 100% bonus on the first top-up (minimum $50). See all plans on the pricing page or download the iRender GPU app.

