Blender Cycles Out of GPU Memory (2026): What Fills VRAM and How to Fit a Scene, Including the 5.2 Texture Cache
Cycles keeps geometry, textures and render buffers in each GPU’s own memory. When a card fills up, CUDA, OptiX, HIP and Metal fall back to system memory and the render slows down. In Blender 5.2 the Texture Cache loads only the texture tiles and resolutions a frame needs, which is usually the biggest saving on texture-heavy scenes.
- Measure: render one frame with -- --cycles-print-stats and compare peak memory with the card.
- Textures: enable Render › Performance › Texture Cache and Auto Generate (Blender 5.2).
- Buffers: lower Tile Size to 1024 or 512 px when memory is tight.
- Scene: Simplify texture resolution and culling, then a larger card if it still does not fit.
A Cycles render that runs out of GPU memory rarely stops outright. Blender moves what does not fit into system memory and keeps going, only more slowly, so the first sign is often a frame that takes far longer than the one before it. Knowing what fills the card tells you which setting to change.
This guide covers what uses GPU memory in Cycles, what happens when a card is full, how the Blender 5.2 Texture Cache works, and the other settings that reduce memory, with their costs.
Checked against the Blender Manual (GPU Rendering, Performance, Simplify, Command Line Arguments) and the Cycles Texture Cache article on code.blender.org, November 2026.
1. What fills GPU memory in Cycles
Three things take most of the space: geometry with its acceleration structure (BVH), image textures, and the render buffers for each pass. Textures grow fastest. The Blender Manual gives an 8K image texture as 256 MB, a 4K texture as 64 MB, a 2K texture as 16 MB and a 1K texture as 4 MB, so a set of large maps adds up quickly (Blender Manual: Image Texture node).
Each GPU needs all of it on its own card. The Manual explains that a GPU can typically only use its own memory, so a second card adds speed, not room; our guide to Cycles multi-GPU rendering covers that in detail.
2. What happens when the scene does not fit
The Manual states that with CUDA, OptiX, HIP and Metal devices, Blender automatically tries to use system memory when GPU memory is full. This has a performance impact, though it usually still renders faster than the CPU. If even that is not enough, the render stops with “Out of memory”, which the Manual describes as not enough memory to store the scene for use by the GPU (Blender Manual: GPU Rendering).
Measure before changing anything. Render one frame from the command line with -- --cycles-print-stats, which the Manual describes as logging render memory and time usage, and compare the peak with your card: 24 GB on an RTX 4090, 32 GB on an RTX 5090 (Blender Manual: Command Line Arguments).
In the middle case the frame still completes, but every access to data held in system RAM crosses PCIe, which is why renders slow down before they fail.
3. The Blender 5.2 Texture Cache
Blender 5.2 LTS added a texture cache to Cycles that loads only the image tiles and resolutions needed (code.blender.org: Cycles Texture Cache). It uses .tx files, which store each image in a tiled, mipmapped format; by default they are written to a blender_tx folder next to the image, and the location can be changed with the Texture Cache file path preference. The Manual describes the trade-off as lower memory use and startup time on scenes with many image textures, at the cost of a small rendering performance impact and more disk space (Blender Manual: Performance).
To turn it on, enable Performance › Texture Cache and Auto Generate in the Render properties. The panel also has a button to generate .tx files for all images used by materials, and one to delete the generated files. Lowering the Tile Size reduces texture cache memory as well, because only the textures needed by each tile are loaded.
On a render farm or another machine, generate the .tx files before you upload, so every node does not build them again. The Manual lists command-line options for this:
Keep the blender_tx folders inside the project you upload, next to the images, so the paths still match. The Manual notes that standard .tx files made by other renderers or OpenImageIO’s maketx are also recognized.
4. Other settings that reduce GPU memory
Tile Size. High-resolution images render in tiles that are cached to disk. The Manual calls the default 2048 pixels a good balance and suggests 1024 or 512 pixels when running out of memory, typically at a small performance cost.
Simplify. In Blender 5.2, the Simplify panel’s Texture Limit was replaced by a Texture Resolution percentage that works with and without the texture cache (code.blender.org). Camera and distance culling remove objects outside the view or far from the camera (Blender Manual: Simplify).
Performance presets and Persistent Data. The Performance panel offers a Lower Memory preset that trades speed for memory. Persistent Data keeps render data in memory between renders for faster re-renders and animations, at the cost of extra memory.
Tiles help because the render buffers and, with the texture cache, the loaded texture tiles only need to cover the part of the frame being rendered. The table below compares these settings with what each one costs.
Settings that lower GPU memory, and what they cost
From the Blender Manual Performance and Simplify pages and the Cycles Texture Cache article.
| Setting | What it reduces | Cost |
|---|---|---|
| Texture Cache + Auto Generate (5.2) | Image texture memory: only needed tiles and resolutions load | Small render-time impact, extra disk space for .tx files |
| Tile Size (Performance › Memory) | Render pass memory, and texture cache memory per tile | Usually a small performance cost at 1024 or 512 px |
| Simplify › Texture Resolution % (5.2) | Texture size across the scene, with or without the cache | Lower texture detail if set too low |
| Simplify › Camera / Distance Culling | Objects outside the view or far from the camera | Missing reflections or shadows from culled objects if margins are too tight |
| Performance preset: Lower Memory | Uses memory-saving settings | Slower renders |
| Persistent Data off | Render data kept in memory between renders | Slower re-renders and animation start-up |
5. When a larger card is the answer
If the measured peak is still close to the card after these changes, a card with more memory keeps the whole scene on the GPU. NVIDIA lists the GeForce RTX 5090 with 32 GB, against 24 GB on the RTX 4090. Our guide to choosing a GPU for Blender covers VRAM, speed and multi-GPU together.
A memory checklist before a heavy Cycles render
- Measure: peak memory from –cycles-print-stats on the heaviest frame
- Texture Cache: enabled with Auto Generate on texture-heavy scenes (Blender 5.2)
- .tx files: generated before upload and kept in blender_tx next to the images
- Tile Size: 1024 or 512 px if memory is tight
- Simplify: texture resolution and culling checked for the final render
- Card: peak below 24 GB or 32 GB with headroom
For paths, caches and other settings to check before any farm render, see our 10-point Blender Cycles pre-flight checklist.
On a submission (SaaS) farm, ask which GPUs and how much memory the render nodes have, and whether they use your .tx files. On an IaaS server you open the scene yourself and watch memory on the exact card you rent.
Checking memory on an iRender server
- Upload the project folder, including blender_tx folders, to iRender Drive before you start the server; uploading is free and does not use rental time.
- Start a server with RTX 5090 (32 GB) or RTX 4090 (24 GB) GPUs and connect by remote desktop.
- Install your Blender version once; the server image is saved between sessions.
- Render one frame with –cycles-print-stats, adjust Texture Cache and Tile Size, then render the sequence and shut down; multi-GPU servers are billed per minute.
Frequently asked questions
1. What happens when a Cycles scene does not fit in GPU memory?
With CUDA, OptiX, HIP and Metal, Blender automatically tries to use system memory. The render continues, but more slowly.
2. What is the blender_tx folder?
It holds the .tx files made by the Cycles Texture Cache in Blender 5.2: tiled, mipmapped copies of your images so Cycles can load only the tiles and resolutions a frame needs.
3. How do I enable the Texture Cache in Blender 5.2?
In Render Properties › Performance, enable Texture Cache and Auto Generate. You can also generate .tx files from the command line with blender scene.blend –command maketx.
4. Does the Texture Cache slow down rendering?
The Blender Manual describes a small rendering performance impact and more disk space in exchange for lower memory use and startup time.
5. Does lowering Tile Size save GPU memory?
Yes. Tiles are cached to disk, so smaller tiles need less render pass memory, and with the texture cache only the textures needed by each tile are loaded.
6. Do two GPUs give Cycles more memory?
Typically no. Each GPU can only access its own memory, so every card must hold the scene.
Render your Blender projects on iRender
iRender is an IaaS GPU render farm: you rent a 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. You install Blender once; your server image is saved between sessions, so Blender and your add-ons stay installed.
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.
This quick start shows how to start a server and render Blender on it:
Two Blender runs we recorded on our servers: 120 frames at 1080p on 4x RTX 5090 at 72 seconds per frame, and the same 4K scene on 4x RTX 5090 and 4x RTX 4090 at 87 and 108 seconds per frame per GPU:
New users get a 100% bonus on the first top-up (minimum $50). See plans on the Blender render farm page or download the iRender GPU app.




