CUDA or OptiX on the RTX 5090: Which Cycles Backend to Choose

Instead of beating around the bush with a long intro: on an RTX 5090, OptiX is almost always the right default choice for Blender Cycles. The main reason is simple: OptiX can use the dedicated RT cores on the RTX 5090, while CUDA runs on the regular CUDA cores. That does not mean CUDA has no place. The result can change with the scene, Blender version, and render settings. With the RTX 5090 and its new Blackwell architecture, it is worth testing both backends on your own projects instead of relying only on older benchmarks. For most heavy Cycles path tracing scenes on an RTX 5090, start with OptiX. If you see unusual results or need to compare output and render time, test the same frame with CUDA as well. One important requirement comes first. Blender 4.4 added support for NVIDIA RTX 50 series GPUs in Cycles, so make sure you are using Blender 4.4 or newer.

In this article, iRender dives deep into how CUDA and OptiX perform on the RTX 5090 to help you optimize your Blender Cycles pipeline.

For an RTX 5090, OptiX is usually the better starting point for Cycles rendering. The reason is the hardware behind the two backends. CUDA uses the GPU’s CUDA cores for general computing. OptiX can use NVIDIA’s hardware ray tracing acceleration, which is designed to speed up ray tracing operations. Blender’s documentation also describes OptiX as using hardware ray tracing acceleration on RTX GPUs.

The RTX 5090 is especially interesting because it uses NVIDIA’s Blackwell architecture. It has 21,760 CUDA cores, 32GB of GDDR7 memory, and fourth-generation RT cores. So if your Cycles scene uses heavy path tracing, OptiX is the first option I would test. But I would not choose a backend based only on the GPU specification. Your own scene is a much better test.

CUDA stands for Compute Unified Device Architecture. It is NVIDIA’s platform for running computing workloads on NVIDIA GPUs. In Blender Cycles, CUDA uses the CUDA cores of the NVIDIA GPU to perform rendering calculations.

It is supported on Windows and Linux. In Blender, you can select CUDA from the Cycles Render Devices settings and then enable the GPU you want to use. For example, if an important production scene gives you a result that looks different from what you expect, rendering the same frame with CUDA gives you another result to compare.

OptiX is NVIDIA’s ray tracing API and is designed to take advantage of dedicated ray tracing hardware on RTX GPUs. This is the main difference that matters when you compare CUDA and OptiX on an RTX 5090. CUDA uses the GPU’s CUDA cores. OptiX can use the RT cores for hardware accelerated ray tracing. Blender’s current documentation specifically notes that OptiX takes advantage of hardware ray tracing acceleration on RTX graphics cards. This makes OptiX a strong choice for Cycles scenes where ray tracing takes a large part of the rendering work. 

The RTX 5090 is based on NVIDIA’s Blackwell architecture and includes fourth-generation RT cores and fifth-generation Tensor cores. It also has 21,760 CUDA cores and 32GB of GDDR7 memory. This matters because OptiX can make use of the RTX hardware designed for ray tracing.

The RTX 5090 is also supported by Cycles starting with Blender 4.4. Blender’s 4.4 release notes specifically added support for the GeForce RTX 50×0 series based on Blackwell. Because the RTX 5090 is a new GPU generation, I would not copy an old CUDA vs OptiX benchmark from an RTX 3090 or RTX 4090 and assume the same result. The better approach is to measure the difference on the actual RTX 5090 and, more importantly, on the scenes you really use. 

CUDA and OptiX are render backends. OptiX Denoiser and OpenImageDenoise are denoisers. They are separate settings. For example, you can render a scene using OptiX and then choose a denoising method separately. You do not have to treat “OptiX” as one single setting that controls everything.

Blender 4.4 also updated the NVIDIA OptiX denoiser and improved its denoising quality. Blender notes improvements in areas such as color accuracy, detail retention, and reduced denoising blotches, although results can still vary between scenes.

OpenImageDenoise is another option. It can also use GPU acceleration on supported NVIDIA RTX GPUs, so it should not be described as a CPU-only denoiser.

Even if OptiX is your first choice, I would still keep CUDA available. There are a few good reasons to test it. One is troubleshooting. If a scene produces an unexpected result, rendering the same frame with CUDA can help you compare the output.

Another is compatibility. If you are using a specific shader, scene setup, add-on, or production workflow that you have not tested on the RTX 5090 before, a second backend gives you another reference point.

The important point is to avoid making broad claims about feature differences between CUDA and OptiX without checking the Blender version you are using. Blender has changed Cycles support over time, so an old comparison may no longer describe the current version.

Because scene complexity directly impacts GPU performance, testing your own files is the best way to determine the right backend for your project.

Follow this simple, step-by-step benchmark workflow:

Step 1: Configure OptiX

Open your scene in Blender 4.4+ and navigate to Edit > Preferences > System. Select the OptiX tab, ensure your RTX 5090 is checked, and render a representative hero frame. Record the exact render time.

Step 2: Configure CUDA

Switch to the CUDA tab in the same System preferences menu. Keep all scene settings identical—including resolution, sample counts, and light bounces—and render the exact same frame. Record the time.

Step 3: Compare Results

Compare the total render times between both backends. Perform a side-by-side visual check on intricate scene details like hair strands, foliage, or glass reflections to ensure consistency.

Step 4: Test Different Scene Types

Repeat this test on both a simple motion graphics file and a heavy architectural or volumetric scene to establish clear pipeline guidelines.

Running benchmark tests sounds simple, but getting physical access to an RTX 5090 just to run a quick test on your local rig represents a massive upfront investment.

This is where an Infrastructure-as-a-Service (IaaS) provider like iRender fits into your workflow. Instead of relying on guesswork or generic online charts, you can rent a remote GPU Cloud Workstation (such as the 3i, 4i or 5i package featuring an RTX 5090 with 32GB VRAM and an AMD Threadripper Pro CPU) for a few hours. 

Let’s see all server’s configurations below: 

Let’s watch the tutorial video to see how our service works:  

1. Is OptiX faster than CUDA on the RTX 5090?

For the vast majority of path-tracing scenes in Blender Cycles, OptiX is significantly faster than CUDA on the RTX 5090. OptiX leverages fourth-generation hardware RT cores for ray-intersection calculations, whereas CUDA relies on standard CUDA cores. Because the ray-tracing hardware on the Blackwell architecture is substantially upgraded, testing OptiX on your specific scene is highly recommended.

2. Do I need a specific Blender version to use OptiX on the RTX 5090?

Yes, you must use Blender 4.4 or newer. Blender 4.4 is the baseline version that introduced full Cycles support for NVIDIA RTX 50-series GPUs based on the Blackwell architecture. Older versions of Blender will not recognize the card correctly for GPU rendering.

3. Is the OptiX denoiser the same thing as the OptiX render backend?

No, they are separate settings in Blender Cycles. The render backend (CUDA or OptiX) controls how light rays and geometry intersections are calculated during rendering. The denoiser (OptiX Denoiser or OpenImageDenoise) is a post-processing step used to clean up noise from the image. Note that OpenImageDenoise has natively supported GPU acceleration since Blender 4.0.

Maximum Speed – Absolute Freedom