If it disappears, increase the value slowly until you find the "sweet spot" before the thread reduction kicks in. 3. Update GPU Drivers
“How much slower?”
The "num_samples_per_thread reduced to 32768" warning is ultimately a sign that your render settings are working against your hardware. By lowering your raw sample counts, turning on Adaptive Sampling, and letting built-in AI denoisers handle the residual grain, you will bypass this hardware limitation entirely and finish your renders in a fraction of the time. If it disappears, increase the value slowly until
However, the warning also speaks to the sophistication of modern error handling. In the early days of computing, exceeding memory limits often resulted in a catastrophic failure: the "Blue Screen of Death" or a silent crash to the desktop. The reduction of samples per thread is an example of graceful degradation. The software sacrifices speed to preserve stability, ensuring that the user eventually gets their image, even if it takes longer. It is a survival mechanism, prioritizing the completion of the task over the efficiency of the process.
A: Absolutely not. num_samples_per_thread is about batching work, not the total samples per pixel. Noise remains unchanged. By lowering your raw sample counts, turning on
Rendering is a task. The ideal scenario: each thread gets a large batch of samples to process, finishes them without interruption, and then moves on. This minimizes scheduling overhead —the time the CPU/GPU spends assigning new work.
If your scene has high max_ray_depth (e.g., 32 bounces), each sample generates many ray states. Multiplying samples per thread by ray depth can quickly overflow memory. The reduction of samples per thread is an
The consequence of this reduction is indicated in the second half of the warning: "rendering might be slower." This slowdown is a result of overhead. When a thread processes fewer samples per cycle, it must loop back to the start of its queue more frequently. This creates "kernel launch overhead" or context-switching costs. Imagine a factory worker who is capable of assembling 100,000 units a day but is only given parts in small baskets of 32,768 units at a time. The worker spends significantly more time walking back and forth to the supply closet (overhead) rather than assembling the product (rendering). The pipeline becomes stuttered, and the raw computational power of the GPU is underutilized because it is constantly waiting for new instructions rather than crunching numbers.