How to render faster in Blender Cycles: 10 settings

Start with the settings that cost nothing in quality: render on the GPU with OptiX, HIP, oneAPI or Metal, enable Noise Threshold with the denoiser, and cap Max Samples. Then cut light bounces to what the scene really needs, clamp indirect light to remove fireflies, turn on Persistent Data for animations, and simplify heavy geometry and textures. Change one setting at a time and time the same frame.

How should you measure render speed before changing anything?

Pick one representative frame, the heaviest of the shot if you can, and render it at final resolution. Blender shows the elapsed time at the top of the render window. Write it down, change a single setting, render the same frame again. Changing five things at once tells you the total gain but not which change broke the glass or darkened the corners.

For quick comparisons you can lower the resolution percentage in Output Properties > Format, but check the final decision at 100%: noise and fireflies behave differently at full size.

1. Is Cycles actually rendering on your GPU?

Open Edit > Preferences > System > Cycles Render Devices and pick the backend for your hardware: OptiX for NVIDIA RTX cards, CUDA for older NVIDIA cards, HIP for AMD, oneAPI for Intel Arc, Metal on Apple Silicon. Tick your graphics card in the list. Then, in Render Properties, set Device to GPU Compute. Both steps are needed: a ticked card with Device left on CPU still renders on the processor.

The CPU can be ticked alongside the GPU. Whether that helps depends on the machine: a fast GPU paired with a modest CPU can end up waiting on the processor. Time it both ways.

2. How many samples does the scene really need?

In Sampling > Render, keep Noise Threshold ticked. Cycles then stops sampling each pixel once it is clean enough, instead of sending the full Max Samples count everywhere. The default of 0.01 is conservative; denoised stills often hold up at 0.02 to 0.05. Max Samples (4096 by default) only matters for pixels that never converge, so lowering it caps the cost of those areas. The noise threshold guide explains how to choose the value on your own scene.

3. Which denoiser should you use?

Tick Denoise in the same panel. OpenImageDenoise works on any hardware and generally preserves texture detail better; OptiX is fast on NVIDIA GPUs and handy in the viewport. Keep Passes on Albedo and Normal. The denoiser is what allows a higher noise threshold, but it cannot rescue a very noisy image without smearing fine detail.

4. How many light bounces do you need?

Light Paths > Max Bounces defaults to Total 12, Diffuse 4, Glossy 4, Transmission 12, Volume 0 and Transparent 8. Every extra bounce costs time for less and less visible light. Lower them, then check what breaks:

Diffuse: exteriors and product shots often look the same at 2. Interiors lit through windows need more, or the rooms darken.
Glossy: 2 or 3 is enough unless mirrors face each other.
Transmission: glass with several layers (a glass behind a glass, a bottle with liquid) turns black when this is too low.
Transparent: leaves and hair cards with alpha cutouts need enough transparent bounces, or black patches appear where they overlap.

5. Should you clamp the light?

Light Paths > Clamping limits how bright a single sample can be. Direct Light is off (0) by default and Indirect Light is set to 10. Lowering Indirect Light, to somewhere between 1 and 5, removes most fireflies, and a frame without fireflies converges sooner under adaptive sampling. The cost is energy: very bright indirect highlights get slightly dimmer. Clamping direct light is rarely needed.

6. Are caustics worth their cost?

Under Light Paths > Caustics, the Reflective and Refractive checkboxes control light focused through glass or bounced off metal onto other surfaces. They are a common source of fireflies and slow convergence. Unless the shot is about the pattern of light under a glass of water, turn them off. Filter Glossy in the same panel blurs glossy paths slightly to reduce noise, at the cost of sharpness in reflections.

7. What does Persistent Data do for animations?

Performance > Final Render > Persistent Data keeps the scene loaded in memory between frames of an animation, so unchanged objects and textures do not have to be prepared again for every frame. On shots where most of the scene is static, this removes a noticeable part of the per-frame overhead. It uses more memory, so turn it off if a heavy scene starts running out of VRAM.

8. Does tile size still matter?

Much less than it used to. Since Cycles X (Blender 3.0), the GPU renders the whole frame progressively. Performance > Memory > Use Tiling, with a default tile size of 2048, exists mainly to keep memory in check on very large renders. Old advice about 256 or 512 pixel tiles for GPUs dates from Blender 2.9 and earlier.

9. How do you lighten a heavy scene?

The Simplify panel in Render Properties has a Render section with Max Subdivision, Child Particles and Texture Limit. They cap subdivision levels, particle children and texture resolution for the whole scene without editing any object, which is a quick way to fit a scene in memory. Camera culling, under Simplify > Culling and enabled per object in its Visibility settings, skips objects outside the camera view. For scenes with many lights, keep Light Tree enabled in Sampling > Lights: it helps Cycles pick the lights that matter for each point, which reduces noise.

10. Why render from the command line?

Rendering in background mode skips the interface and the viewport, which leaves more memory for the render itself. From a terminal:

blender -b shot.blend -a              # render the whole animation
blender -b shot.blend -s 1 -e 120 -a   # frames 1 to 120 only
blender -b shot.blend -f 42            # a single frame

Options are read in order, so -s and -e must come before -a. The step by step for a full animation, output format included, is in the guide to rendering an animation in Blender.

Frequently asked questions

Why is my GPU render slower than my CPU render in Blender?

The usual causes are a scene that does not fit in the graphics card memory, CUDA selected on an RTX card instead of OptiX, or the GPU device not ticked in Preferences > System, which leaves Cycles on the CPU. Check the device list first, then watch memory use during the render.

Is OptiX faster than CUDA?

On NVIDIA RTX cards, OptiX uses the ray tracing cores and is usually the faster of the two. CUDA remains useful on older NVIDIA cards without those cores. Render the same frame with each to see the difference on your scene.

Does a smaller tile size make Cycles render faster?

Not since Cycles X in Blender 3.0. The GPU renders the whole frame progressively, and tiling now exists mainly to reduce memory use on very large images. Leave the default tile size unless you run out of memory.

Should I switch to Eevee to render faster?

Eevee is a rasterising engine and usually renders a frame much faster than Cycles, but it approximates reflections, refractions and indirect light differently. It is a change of look, not a speed setting. Switch only if the result is acceptable for the shot.

Rendering the animation on TURNA

When a scene is already well optimised and the frame count is what takes the time, the remaining lever is rendering several frames at once. TURNA does this for Blender: an add-on (Blender 4.0 to 5.2, Cycles and Eevee) sends the scene in one click, and the frames are split across 1, 2, 4 or 8 GPUs in parallel, RTX 5090 first and RTX 4090 when none is free. The add-on shows the estimated cost before anything starts. Prices start at €2.90 per hour, billed on the time actually used, with a 30-second minimum, and new accounts get €25 of free credit. A single still image is still rendered by one machine, so the gain is on animations. Tiers and worked examples are on the Blender render farm page.

How to optimize Cycles to render faster in Blender