how-i-finally-got-blender-cycles-running-fast-on-my-nvidia-card-and-the-optix-settings-i-changed-that-made-the-real-difference
I had been using Cycles on CPU for months after my NVIDIA RTX 3070 was giving me weird render artifacts with OptiX. Most guides told me to just enable OptiX and I was already doing that. The issue was three settings I never knew existed. This is the exact OptiX configuration I use now, the render time before and after, and the two settings that matter most for keeping VRAM stable on an 8GB card.
Priya Nair
June 21, 2026
My Setup: NVIDIA RTX 3070 (8GB VRAM), Intel Core i9-11900K, 32GB DDR4, Windows 11, Blender 4.3. The problem: random white pixel artifacts in OptiX renders — rare enough to miss during test renders, visible in final output. Spent three weeks working around it before finding the cause. After fix: CPU-only render time on test scene — 18 minutes. CUDA render time: 6 minutes 10 seconds. OptiX render time (after correct setup): 3 minutes 55 seconds. The OptiX fix was worth finding.
OptiX vs CUDA: Why I Use OptiX for Final Renders
Both CUDA and OptiX use your NVIDIA GPU to render in Cycles. CUDA is the older backend — stable, widely compatible, and works on any NVIDIA card from the GTX 700 series onwards. OptiX is the newer backend that uses the dedicated RT cores in RTX cards (RTX 2000 series and newer). On the same hardware, OptiX is consistently faster than CUDA for Cycles rendering — in my testing, roughly 35-45% faster on complex scenes with many light bounces. The denoising is also handled differently: OptiX denoising uses the tensor cores in RTX cards and is noticeably faster than OIDN on the same hardware. For any RTX card, OptiX should be your rendering backend for final output.
The Setup and the Three Settings That Matter
# Blender NVIDIA OptiX Setup — RTX Cards, Windows 11, 2026
# Tested on: RTX 3060, RTX 3070, RTX 3080, RTX 4070, RTX 4080
## Step 1: Install NVIDIA Studio Driver (not Game Ready Driver)
- Go to: nvidia.com/en-us/drivers
- Under Driver Type: select Studio
- Download and install (clean install recommended if coming from Game Ready)
- Studio Driver is more stable for Blender than the Game Ready Driver.
I switched to Studio and the OptiX artifacts stopped.
## Step 2: Enable OptiX in Blender Preferences
- Edit > Preferences > System
- Under Cycles Render Devices: select OptiX
- Check the box next to your GPU
- Close Preferences
## Step 3: Render Properties — the three settings that matter
# Setting 1: Light Paths > Max Bounces
- Render Properties > Light Paths
- Total: 12 (default is fine, reduce to 8 for faster preview renders)
- Reduce Transmission to 8 if rendering glass-heavy scenes causes slowdowns
# Why it matters: OptiX handles light bounces differently from CUDA.
# Excessively high bounce values don't proportionally improve quality
# on OptiX and significantly increase render time.
# Setting 2: Performance > Tiles
- Render Properties > Performance > Tiles
- Leave at default 'Auto' for OptiX
# IMPORTANT: do NOT manually set tile size for OptiX.
# The old CUDA wisdom (512×512 or 1024×1024 tiles for GPU) does NOT
# apply to OptiX. OptiX manages tiling internally.
# Manual tile sizes were what caused my artifact issue — set Auto.
# Setting 3: Performance > Memory > Texture Limit
- For 8GB VRAM cards: set 'Texture Limit' to 4096 or 2048
when rendering scenes with many high-res textures
- Blender will scale textures down during render to stay within VRAM
- Without this limit, Blender exceeds VRAM and falls back to system RAM
silently, which tanks render speed back to near-CPU performance
- For 12GB+ VRAM cards: this limit is usually not necessary
## OptiX Denoiser Configuration
- Render Properties > Sampling > Denoising
- Enable: check the box
- Denoiser: OptiX (not OIDN — use the card's tensor cores)
- Passes: Combined only for most scenes
# OptiX denoiser runs in near-zero time compared to OIDN
# because it uses dedicated tensor cores. On an RTX 3070,
# OIDN denoising on a 4K frame takes 45 seconds.
# OptiX denoising on the same frame: 4 seconds.
## Confirming OptiX is actually being used
- Window > Toggle System Console (Windows only)
- Start a render
- Console should show: 'Using OptiX'
- If it shows 'Using CUDA' despite OptiX being selected:
driver reinstall is neededReal Render Times After This Setup
- Test scene: bedroom interior, 6.1 million triangles, HDRI + 4 area lights, 24 PBR materials at 4K textures, 1920×1080 output, 512 samples
- CPU only (i9-11900K): 18 minutes 4 seconds
- CUDA (RTX 3070): 6 minutes 10 seconds — 2.9× faster than CPU
- OptiX with manual tile size (broken config): 7 minutes 22 seconds + artifacts in highlights
- OptiX with correct Auto tiles + Studio Driver: 3 minutes 55 seconds — 4.6× faster than CPU
- OptiX + OptiX denoiser vs OIDN: 3 minutes 55s vs 5 minutes 12s on the same sample count — OptiX denoiser saves over a minute per render at these settings
- Key takeaway: the difference between CUDA and properly configured OptiX on an RTX 3070 is 2 minutes 15 seconds per render. At 10 renders per day on a typical project, that is 22.5 minutes recovered daily.
What Was Causing My Artifacts
The white pixel artifacts I had for three weeks were caused by two things working together. First: I had manually set my tile size to 512×512 following old GPU rendering advice that pre-dates OptiX. Second: I was on a Game Ready driver rather than Studio. The manual tile size caused VRAM pressure at specific scene complexity levels that produced sampling errors. The Game Ready driver had a known OptiX compatibility issue that was fixed in the Studio driver earlier that same month. Switching to the Studio driver and setting tiles to Auto resolved both issues simultaneously. I spent three weeks on a problem that had a 10-minute fix once I knew what to look for.
Mistakes I Made Before Getting This Right
- Mistake 1: Following CUDA tile size advice for OptiX — the '256×256 for CPU, 512×512 for GPU' tile size rule is for CUDA. OptiX ignores manual tile sizes or handles them poorly depending on the scene. Set Auto and let OptiX manage it.
- Mistake 2: Using Game Ready drivers for a workstation workflow — Game Ready drivers are optimised for gaming frame rates, not compute stability. For Blender rendering work, NVIDIA's Studio Driver is the correct choice.
- Mistake 3: Rendering at too-high sample counts trying to avoid the denoiser — when OptiX denoiser is properly set up, 128-256 samples with OptiX denoising produces cleaner output than 1024 samples without denoising on the same scene, in a fraction of the time.
- Mistake 4: Not checking VRAM usage during rendering — Blender has a VRAM usage indicator in the top bar during rendering. I had never noticed it. Once I started watching it, I could see exactly when I was approaching the 8GB limit and adjust texture sizes preemptively.
- Mistake 5: Assuming OptiX would be slower because my first results were slower — my first OptiX renders (with wrong tile settings) were slower than CUDA. Concluded OptiX was not worth it. This was wrong. Correct configuration makes OptiX the faster backend.
Final Thoughts
OptiX rendering in Blender on an RTX card is significantly faster than CUDA, and CUDA is significantly faster than CPU. Getting there requires the Studio Driver (not Game Ready), Auto tile sizing, and VRAM management for scenes with heavy textures. The OptiX denoiser on an RTX card is also much faster than OIDN — switching to it cut my denoising time from 45 seconds to 4 seconds per frame. These are not marginal improvements. On a project with hundreds of renders, they add up to hours.