OptiX vs CUDA in Blender: choosing and setting up your NVIDIA card

An NVIDIA card does not render Cycles until you ask it to: by default, Blender renders on the CPU. This guide explains where to choose between CUDA and OptiX, what OptiX actually speeds up, which cards are concerned and the known limits, based on the Blender 5.2 LTS manual and the official release notes.

CUDA and OptiX: what are they?

CUDA is NVIDIA's general-purpose computing platform. In Blender, it is one of the Cycles "backends": the engine runs on the graphics card like any ordinary compute program.

OptiX is NVIDIA's ray tracing technology. As a Cycles backend, it does the same job as CUDA, rendering the image on the card, but it hands part of the work to the dedicated hardware of RTX cards. The Blender manual puts it this way: OptiX takes advantage of the hardware ray tracing acceleration of RTX cards for better performance.

The same name also refers to a denoiser: the OptiX denoiser, based on artificial intelligence, which Cycles offers alongside OpenImageDenoise. The two settings are independent, so you can render with the OptiX backend and denoise with OpenImageDenoise.

Finally, this choice only concerns Cycles. EEVEE is a rasterisation engine, not a path tracer: it draws the scene through the graphics driver (OpenGL or Vulkan) and uses neither CUDA nor OptiX.

Where to set CUDA or OptiX, so that Blender uses the GPU

1. Open Edit > Preferences, then the System section.
2. Under Cycles Render Devices, choose the OptiX or CUDA tab. The HIP, oneAPI and Metal tabs are for AMD cards, Intel cards and Macs with Apple silicon. If only None is offered, Blender found no usable card and will render on the CPU.
3. Tick the card or cards to use. Depending on the backend, the CPU can also be ticked to take part in the render.
4. In each scene, open Render Properties, choose Cycles and switch Device from CPU to GPU Compute. This setting is saved in the .blend file, so a file that comes from another computer may still be set to CPU.

From the command line, the --cycles-device option overrides the preferences for a render. It goes after a double dash, for example blender -b scene.blend -f 1 -- --cycles-device OPTIX. The accepted values are CPU, CUDA, OPTIX, HIP, ONEAPI and METAL; adding +CPU (OPTIX+CPU) makes the CPU work alongside the card.

What OptiX speeds up

Rendering with Cycles means casting rays and finding, for each one, the surface it hits. To avoid testing every triangle in the scene, the engine stores the geometry in an acceleration structure, the BVH, a tree of nested bounding boxes. Traversing this tree and testing intersections with triangles takes up a large share of the render time.

RTX cards have RT cores, units that NVIDIA designed for exactly this work: BVH traversal and ray-triangle intersection tests run in hardware instead of keeping the ordinary compute cores busy. With CUDA, Cycles does this traversal in software on the CUDA cores; with OptiX, it hands it to the RT cores. The rest of the work, evaluating materials and lighting, stays on the ordinary compute cores in both cases.

The gain therefore depends on the scene: the more the intersection search weighs in the render time, the larger it is. The manual gives no factor and refers to Blender Open Data, Blender's public database of measurements, to compare cards and backends. The Blender benchmark page (in French) reproduces its ranking, backend included.

The OptiX denoiser

In Render Properties, under the Sampling panel, the Denoise section offers three denoisers:

• Automatic uses GPU-accelerated denoising when it is available, preferring OpenImageDenoise to OptiX;
• OpenImageDenoise, Intel's denoiser, which the manual describes as the one that generally gives the best quality;
• OptiX, NVIDIA's denoiser, available only on NVIDIA cards configured in Cycles Render Devices. It stays GPU-accelerated on some older models where OpenImageDenoise does not.

OpenImageDenoise runs on the graphics card from compute capability 7.0, which includes every RTX card. The Use GPU checkbox turns this on: it is much faster than on the CPU but uses more video memory, so untick it if a large scene runs out of memory. In both cases, give the denoiser at least the Albedo pass, and preferably Albedo and Normal: without them, it may smooth away details. The OptiX denoiser was improved in Blender 4.4, then again in Blender 5.0.

Which NVIDIA cards are concerned?

CUDA and OptiX work on Windows and Linux, not on Mac. Both need a card with compute capability 5.0 or higher, which roughly means the GeForce 900 series and later; Blender 5.0 dropped older cards.

On a GTX card, OptiX works but has no RT cores to use: render the same frame with both backends and keep the faster one. A very recent card also needs a version of Blender that knows it: GeForce RTX 50 cards have been supported since Blender 4.4. With an older version, Cycles has no precompiled kernel for the card and can fail with the error "Kernel compilation failed"; the manual then advises moving to a more recent version of Blender.

The NVIDIA driver

OptiX depends on the graphics driver. The required minimum was raised to version 535 in Blender 4.5, then to version 575 in Blender 5.2. The manual recommends the official drivers from the graphics card manufacturer, since the ones from the computer manufacturer are sometimes old or incomplete. If the OptiX tab disappears after a Blender update, check the driver version first.

Known limits

• Video memory. A card only sees its own memory. When it is full, CUDA and OptiX spill over into the computer's memory, which slows the render down while usually staying faster than the CPU. Two cards do not add up their memory, except NVIDIA cards linked by NVLink with the Distribute Memory Across Devices option. The manual gives a rule of thumb: an 8K texture takes 256 MB, a 4K one 64 MB.
• Path Guiding, which reduces noise in scenes lit through a small opening, is not available on any graphics card: it only works on the CPU.
• Open Shading Language works on the CPU and, among the GPU backends, only with OptiX, with a few limitations. This is a case where OptiX is required, not just faster.
• Interface freezing during the render. A card that renders and drives the display at the same time can make Blender sluggish. On Windows, a long render can even exceed the time the system allows the display card; reducing the tile size (Performance panel) helps, but the real fix is a card dedicated to rendering.

OptiX or CUDA: which should you choose?

• RTX card: OptiX. It is the only backend that uses the RT cores, and it gives access to Open Shading Language on the GPU.
• Recent GTX card: both work; measure on your own scene.
• Rendering from the command line or on a server: set the backend with --cycles-device rather than relying on the machine's preferences. If the requested backend does not exist, Blender reports it with a message such as "Found no Cycles device of the specified type".

If the render is still slow despite the card, the cause is often the settings rather than the backend: samples, noise threshold or bounces. The noise threshold guide and the guide to rendering faster in Cycles go through these levers.

How TURNA uses OptiX

On TURNA, Cycles renders run with OptiX by default, through the --cycles-device OPTIX option, on RTX cards (RTX 5090 first, RTX 4090 otherwise). If the compute image does not allow this option to be set, the node falls back to CUDA and records it in its log. The denoiser chosen in your scene is not replaced: TURNA does not swap OpenImageDenoise for OptiX, since the two do not produce exactly the same image, but it runs your denoiser on the graphics card when Blender allows it. A machine on which OptiX fails to initialise is classed as a hardware failure, not as an error in your scene. EEVEE renders do not use OptiX.