System Information
Operating system: Windows 10 1909 or Linux Ubuntu 19.10 (same issue with either operating system)
Graphics card: multiple Nvidia GTX 1060 3GB or 6GB
Blender Version
Broken: 2.82a release
Worked: None that I know of
Short description of error
When rendering with Cycles and CUDA with GPU compute and only GPUs, CPU utilization is high and increases proportionate to the number of GPUs added. This is only observed on older Intel systems so far such as 3rd gen and 6th gen Intel Core i5 systems/boards. On those systems the CPU will hit 100% after ~3 GPUs are added when GPU-only rendering. On a Xeon Platinum system, 7 GPUs barely require 12% of CPU power. I have not yet tested this issue on Ryzen.
Exact steps for others to reproduce the error
Found using the Classroom .blend file (current version) from the Blender website, but occurs with other scenes
- Use an affected system such as Intel H110 chipset motherboard and 6th-gen Intel Core i5 CPU (e.g. 6502P) with more than one Nvidia GPU (e.g. GTX 1060, but also older cards like 760)
- Software: Windows 10 x64 1909 or Ubuntu 19.10 or OpenSUSE Tumbleweed with Nvidia 440.xx driver (including CUDA) and Blender 2.82a
- In Blender Preferences -> System, choose CUDA for rendering device and select only one Nvidia GPU and no CPU.
- In render settings, set to Cycles, GPU compute. Tile size, samples, and other settings can be left at the default of the blend file
- Hit that F12 key and start rendering
- Review top, htop, Task Manager, or whatever tool you want to view CPU use.
- Observe that when rendering with one GPU, the CPU is using more than one core's worth of CPU performance (37% in my case).
- Return to Blender Preferences -> System and enable a second GPU under CUDA (still no CPU).
- Hit F12 again
- Observe CPU utilization is now ~2x as high as a single GPU (~60%)
- Repeat for third GPU, and observe GPU utilization is now almost 100%
Other notes
- Per GPU-z and nvidia-smi, the GPUs were not out of VRAM. The Classroom scene rendered in Cycles uses 1.22GB of VRAM. All GPUs I used had 3GB or 6GB. Similarly, the system never exceeded 65% of system RAM used.
- More discussion can be found in this thread