April 10, 2025 Jun Do

Why Is Blender Rendering So Slow? How to Make Cycles and EEVEE Render Faster

Executive SummaryFaster Blender renders at a glance
Blender 4.x and 5.x

Most slow Blender renders are not a weak GPU. Check them in this order: is Cycles really on the GPU, are you brute-forcing noise with samples, does the scene fit in VRAM, and does the image need every light bounce and every subdivision level.

  • GPU first: Preferences › System › Cycles Render Devices (OptiX on NVIDIA RTX), and Render Properties › Device › GPU Compute
  • Samples: Noise Threshold plus a denoiser instead of thousands of samples
  • Scene: smaller and 8-bit textures, instances instead of copies, Simplify for subdivision and texture size
  • Light and animation: Max Bounces to what the shot needs, Light Tree for many lights, Persistent Data for animations
  • Still slow: Cycles splits every frame across all GPUs in the machine; VRAM is not pooled, so the scene must fit on one card

In short: fix the settings first wherever you render, then add GPUs for the frames that are still slow.

“Why does Blender take so long to render?” usually has a dull answer: the GPU is not being used, the sample count is far higher than the image needs, or the scene carries data nobody sees. This guide goes through the causes in the order to check them, from the one-minute fixes to the scene work, and ends with what to do when a heavy scene is still slow.

It brings together our earlier guides on making Cycles render faster, slow Blender renders and reducing render time, updated for Blender 4.x and 5.x.

Checked against the Blender Manual (4.4–5.0) and Blender release notes, November 2026.

1. Is Blender actually rendering on the GPU?

This is the most common reason for a render that suddenly takes forever. Cycles needs two settings, and files downloaded from elsewhere often have the second one set to CPU:

  1. Open Edit › Preferences › System. Under Cycles Render Devices choose the backend for your card: OptiX for NVIDIA RTX (CUDA also works on NVIDIA), HIP for AMD, oneAPI for Intel Arc, Metal on Apple.
  2. Tick every GPU you want to use. Untick the CPU unless you have tested that it helps.
  3. In Render Properties, set Device to GPU Compute. This setting is saved in the .blend file.
  4. Render and check the load: in a Command Prompt nvidia-smi shows GPU-Util per card. In Windows Task Manager, switch a GPU graph to Cuda; the default 3D graph stays near 0% during Cycles renders.
iRender
Making sure Cycles renders on the GPU
Schematic
Edit › Preferences › System › Cycles Render DevicesNoneCUDAOptiXHIPoneAPIMetal1NVIDIA GeForce RTX 5090NVIDIA GeForce RTX 5090AMD Ryzen Threadripper PRO (CPU)2OptiX: NVIDIA RTX; HIP: AMD; oneAPI: Intel Arc;Metal: Apple. Tick every GPU you want to use.Preferences save automatically unless Auto-Save is offRender Properties (Cycles)DeviceGPU Compute3saved per .blend: downloaded files oftencome back set to CPUWhile rendering: check the GPU worksnvidia-smiGPU 0GPU-Util97%GPU 1GPU-Util96%4Near 0% while Blender renders? The scene is on the CPU: go back to badges 1 to 3. Task Manager shows Cuda load, not the 3D graphOrange outline = the setting to change; badge = step number in the textLayout reconstructed from the Blender Manual (Preferences System, GPU Rendering); simplified
Setting to changeTicked / selected
Simplified mockup; example values

How to read the figure. Badges 1 and 2 are in Preferences, which apply to every file: the backend and the cards to use. Badge 3 is saved in each .blend, which is why a file from someone else can render on the CPU while your own files use the GPU. Badge 4 is the check while a frame renders: each GPU should show high utilisation. If it stays near 0%, go back to badge 3 first, then to badges 1 and 2.

2. Stop brute-forcing noise with samples

Cycles noise falls only with the square root of the sample count: four times the samples halve the grain and take about four times as long. So a render at thousands of samples is often slow for little gain.

  • Noise Threshold (Render Properties › Sampling › Render) stops sampling each pixel once it is clean enough. The default is 0.01.
  • Max Samples only needs to be high enough for the darkest, noisiest areas to reach the threshold.
  • Denoise with OpenImageDenoise or OptiX cleans what is left. Keep the Albedo and Normal passes on. For animation, frames denoised one by one can shimmer; see our guide to denoising Cycles animations.

Our guide to fixing grainy renders shows these settings and firefly clamping step by step.

3. Textures: resolution and bit depth

An image takes width × height × channels × bytes per channel in memory. Doubling the resolution quadruples the memory, and a 32-bit float image takes four times as much as an 8-bit one. Keep float images (EXR, HDR) for data that needs them, such as HDRIs and displacement, and use 8-bit images for color textures.

  • Size textures to the screen. A texture on an object that covers a small part of the frame does not need 8K.
  • Cap them for a test with Render Properties › Simplify › Render › Texture Limit (Cycles). If the frame looks the same at 2048 or 4096, resize the images for good.
  • Reuse images. The same image used in several materials is stored once; ten slightly different copies are stored ten times.
iRender
How much VRAM one texture takes
Schematic
Memory of one RGBA image in VRAM = width × height × 4 channels × bytes per channelbar length on a square-root scale so small and large values both stay readable1K (1024²)4 MiB8 MiB16 MiB2K (2048²)16 MiB32 MiB64 MiB4K (4096²)64 MiB128 MiB256 MiB8K (8192²)256 MiB512 MiB1 GiBOne 8K float texture = 1 GiB, as much as 256 8-bit 1K textures (4 MiB each)Every GPU keeps its own copy: VRAM is not pooled across GPUsValues computed from the formula; Blender can also keep some images in compressed or half-float form, so treat these as upper bounds per textureFormula: pixels × channels × bytes; Blender Manual: Cycles GPU Rendering (memory), Simplify (Texture Limit)
8-bit16-bit half32-bit float
Computed values, square-root bar scale

How to read the figure. Each group is one resolution, each bar one bit depth, with values computed from the formula above the chart. Going down a group multiplies memory by four, going from blue to red within a group also multiplies it by four. The bars use a square-root scale so 4 MiB and 1 GiB both fit; the numbers next to them are the real values. Remember that each GPU in a multi-GPU machine holds its own copy, so this memory is needed on every card.

4. Geometry: instance, do not copy

  • Instances. Duplicate repeated objects with Alt+D (Duplicate Linked), collection instances or Geometry Nodes instances. All instances share one mesh; a normal duplicate (Shift+D) stores the mesh again.
  • Subdivision. Subdivision is often the largest hidden cost. Check the Render levels of Subdivision Surface modifiers, keep background objects low, and cap all of them at once with Simplify › Render › Max Subdivision. In Cycles, adaptive subdivision can subdivide only where the camera needs detail.
  • Clean meshes. Merge doubled vertices with Mesh › Clean Up › Merge by Distance and delete hidden geometry that never appears on camera.
iRender
Why instancing saves memory
Schematic
1,000 trees in a scene: real copies against instances, measured in units of one tree meshReal copies1,000 × meshInstances1 × mesh + 1,000 transforms (64 bytes each = 64 KB in total)0×250×500×750×1000×A transform is a 4 × 4 matrix of 32-bit floats: 16 × 4 bytes = 64 bytes. The mesh data is stored once.Linked duplicates (Alt+D), collection instances and Geometry Nodes instances share one mesh in memoryBlender Manual: Duplicate Linked, Collection Instances, Geometry Nodes Instances; matrix size computed
CopiesInstances
Drawn to scale in mesh units

How to read the figure. The red bar is 1,000 separate copies of a tree, so the mesh is stored 1,000 times. The green sliver is 1,000 instances: the mesh is stored once and each instance only adds a 64-byte transform that says where it stands, about 64 KB for all of them. That difference is why forests, crowds and scattered rocks should always be instances.

5. Light paths, lights and materials

  1. Max Subdivision (Simplify): lower it to see whether high subdivision levels are needed in this shot.
  2. Texture Limit (Simplify, Cycles): cap textures to test how much resolution the frame really needs.
  3. Max Bounces (Light Paths): the defaults are Total 12, Diffuse 4, Glossy 4, Transparent 8. Lower the ones your scene does not use much; keep Transparent high enough for stacked leaves or hair, or they turn black. Turn off Reflective and Refractive Caustics if the shot does not need them.
  4. Light Tree (Sampling › Lights): helps scenes with many lights. With only a sun and an HDRI, render a test frame with it on and off.
  5. Persistent Data (Performance › Final Render): for animations, keeps scene data in memory between frames instead of rebuilding it, at the cost of more RAM.

Materials matter too: volumes, subsurface scattering and deep glass are the expensive effects, so keep them where they show. Large procedural node trees can be baked to image textures once the look is final.

iRender
The scene-level render settings, in the order to try them
Schematic
Render Properties › Simplify › RenderMax Subdivision2 (lower)1Child Particles1.000Texture Limit2048 (cap)2tick Simplify in the panel header first; Cycles only for Texture LimitRender Properties › Light Paths › Max BouncesTotal12 → lower3Diffuse4Glossy4Volume0Transp.8Fast GI Approximationdefaults shown; lower only what the scene does not needRender Properties › Sampling › LightsLight Tree4helps scenes with many lights; a scene with one sunand one HDRI may render faster without it — test bothRender Properties › Performance › Final RenderPersistent Data5keeps scene data in memory between frames of ananimation; uses more RAM while it rendersRender the same frame after each change and compare time and image; keep a change only if the image still looks rightOrange outline = the setting to change; badge = step number in the textLayout reconstructed from the Blender Manual (Cycles Simplify, Light Paths, Sampling Lights, Performance); Blender 4.x and 5.x; simplified
Setting to changeOption on
Simplified mockup

How to read the figure. The top row is about data: badge 1 caps subdivision and badge 2 caps texture size, both in the Simplify panel, which must be ticked in its header. The top right is about work per pixel: badge 3, the bounce limits. The bottom row helps specific scenes: badge 4 for many lights, badge 5 for animations. Change one badge at a time and render the same frame, keeping a change only if the image still looks right.

6. Drivers, Windows settings and old advice

  • Driver. Use a current driver from your GPU maker; NVIDIA’s Studio drivers are tested with creative apps. A driver older than your Blender version expects can stop OptiX from working.
  • Hardware-accelerated GPU scheduling. This Windows setting rarely changes Cycles render times; if you suspect it, test a frame with it on and off as described in our HAGS guide.
  • Tiles and Progressive Refine. Advice to tune tile sizes or turn off Progressive Refine is from Blender 2.9x and older. Cycles X in Blender 3.0 removed Progressive Refine and no longer needs small tiles; leave the tiling options at their defaults unless you run out of memory on very large images.
  • Test at lower resolution. Render Properties › Output › Resolution % at 50% renders a quarter of the pixels for quick checks.

7. EEVEE: one GPU, samples and effects

EEVEE renders with one GPU and has no bounces to lower. Its cost is mostly the Render sample count, shadows and ray-tracing settings, plus the same texture and geometry memory as Cycles. Lower Render samples until soft shadows and depth of field start to look grainy, then go one step back up. For long EEVEE animations, split the frame range across several single-GPU machines instead of using one multi-GPU machine.

8. Split heavy scenes and composite

When one part of a shot is expensive, such as a volume, hair or a big environment, render it as its own View Layer and combine the layers in the compositor. You can re-render only the layer that changed. Our guide to Blender View Layers shows the setup.

If Blender shows “Render size too large for GPU, use CPU compositor instead”, the GPU compositor cannot handle an image that large. Switch the compositor’s device to CPU in Render Properties › Performance › Compositor, or render the passes and composite at a smaller size.

9. When the scene is simply heavy

If you have worked through the list and a frame is still slow, the scene may just be heavy. Two facts decide what helps next:

  • Cycles uses every GPU for each frame. With 2, 4 or 8 GPUs ticked, a frame is shared between them, so more GPUs shorten every frame.
  • VRAM is not pooled. Each GPU holds the whole scene, so the largest scene is set by one card: 24 GB on an RTX 4090, 32 GB on an RTX 5090. If Cycles runs out of VRAM it can continue with system RAM, more slowly. Our guide on how much VRAM you need shows how to measure a scene.

Faster Blender render checklist

Tick each item before you render the full sequence.

ChecklistFaster Blender render checklist
0 / 8 done
Tip: tick each item as you set up the job; the ticks reset when the page reloads.

SaaS or IaaS render farm for a slow Blender scene?

Every fix above is saved in the .blend, so it helps on any farm. What a farm changes is the hardware: more GPUs make each frame faster, and a card with more VRAM lets a heavy scene stay in GPU memory. Rendering in the cloud does reduce render time when the frame is limited by GPU power; it does not fix a scene that is set to the CPU or to far too many samples.

A speed-test workflow on an iRender server

iRender
How iRender works: your computer controls your own multi-GPU server
IaaS render farm
Your PC or MacNo powerful GPU neededRemote desktopLive screen1:1 like your PCYOUR OWN GPU SERVERbare-metal · not sharedGPUGPUGPUGPUGPUGPUGPUGPU1, 2, 4 or 8 × RTX 5090 (32 GB) or RTX 4090 (24 GB)Threadripper PRO · 256 GB ECC RAM · 2 TB NVMe SSDFull admin access · your setup is saved between sessions1Register, top upFirst $50+ top-up doubled2Upload your projectiRender Drive, free3Start and connectRemote desktop, any OS4Render, shut downBilling stops, per minute✓ Your workstation, just more GPUs: no queue to wait, no farm plugin, no job submission, no settings changed
Your mouse and keyboardServer screen streamed back
Schematic, not to scale

How to read the figure. The top row is the setup: your own PC or Mac connects by remote desktop to a whole bare-metal server with 1, 2, 4 or 8 GPUs, so you open the .blend there exactly as at home. The numbered row is the order of a session: register and top up, upload to iRender Drive before the server runs, start and connect, render a test frame and time it, render the job, then shut down so billing stops.

  1. 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.
  2. Pick the server. RTX 4090 servers have 24 GB per GPU, RTX 5090 servers 32 GB per GPU. Cycles uses every GPU in the server for each frame; EEVEE uses one.
  3. Install Blender from blender.org (it is not pre-installed); it stays in your server image. Set OptiX and GPU Compute as in section 1.
  4. Measure. Render the heaviest frame and note the time and peak memory.
  5. Render the range. Multi-GPU servers bill per minute; 1-GPU servers have a one-hour minimum. Shut the server down afterwards so billing stops.

Frequently asked questions

1. Why does Blender take so long to render?

Usually Cycles is rendering on the CPU, the sample count is far higher than the image needs, or the scene holds heavy textures and subdivision nobody sees. Check the device first, then samples and denoising, then the scene.

2. Why is Blender not using my GPU?

Two settings are needed: a backend and your GPUs ticked in Edit › Preferences › System › Cycles Render Devices, and Device set to GPU Compute in Render Properties. The second is saved per file.

3. Which Cycles render device should I choose?

OptiX for NVIDIA RTX cards, CUDA for older NVIDIA cards, HIP for AMD, oneAPI for Intel Arc and Metal on Apple computers.

4. How many samples do I need in Blender?

There is no fixed number. Turn on Noise Threshold and Denoise, then raise Max Samples only until the darkest areas stop being grainy.

5. Does Persistent Data make Blender render faster?

For animations, yes: it keeps scene data in memory between frames instead of rebuilding it for every frame. It uses more RAM while it renders.

6. What does Render size too large for GPU, use CPU compositor instead mean?

The GPU compositor cannot process an image that large. Set the compositor device to CPU in Render Properties › Performance › Compositor, or composite at a smaller size.

7. Does cloud rendering reduce render times?

When a frame is limited by GPU power, yes: Cycles splits each frame across every GPU in the machine. It does not help a scene that renders on the CPU or uses far more samples than it needs.

8. Do more GPUs add VRAM in Blender?

No. Each GPU keeps its own copy of the scene, so the largest scene is limited by one card: 24 GB on an RTX 4090, 32 GB on an RTX 5090.

9. Is EEVEE faster than Cycles?

Usually, because EEVEE is a real-time engine. It renders with one GPU, and its render time depends mostly on the sample count, shadows and ray-tracing settings.

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.

Accelerate your production pipeline with high-performance IaaS render farm services and dedicated bare-metal workstations powered by up to 8x NVIDIA GeForce RTX 5090 (32GB GDDR7), high-frequency AMD Ryzen™ Threadripper™ PRO processors, 256GB ECC RAM, and dedicated 2TB NVMe SSD storage tailored for your pipeline:
, , , , , , , , , , , , , , , ,

Jun Do

Learn something new everyday with iRender
Contact

iRENDER GPU

Render Farm Pricing
Quick Start
Multi-GPU Benchmarks
iRender GPU App
Blog
Documentation
The Render Farm
The Company
Terms & Conditions

INTEGRATIONS

Autodesk Maya
Blender
Cinema 4D
Houdini
Karma XPU
Lumion
KeyShot
Unreal Engine
Twinmotion
Redshift
Octane
And many more…

iRENDER TEAM

Elite support 7 days a week
SATURDAY – SUNDAY: 6:00 AM – 11:59 PM
(UTC+7)
Hotline: (+84) 912-785-500
Skype: iRender Support
Email: [email protected]
Address 1: 68 Circular Road #02-01, 049422, Singapore.
Address 2: No.22 Thanh Cong Street, Hanoi, Vietnam.

Contact
이 페이지는 한국어로도 제공됩니다. 한국어로 보기 →