get_picture("Color") takes ~0.03s using integrated CPU graphics — is this normal? What speed can be expected with a dedicated GPU (e.g., RTX 4060)?

[pioneer-innovation]

Hello,
When testing the get_picture("Color") function using only the integrated GPU on the CPU (e.g., Intel UHD Graphics), the average execution time per call is about 0.03 seconds.

I would like to ask:

Is this execution time considered normal when running solely on integrated CPU graphics?
If a dedicated GPU (such as NVIDIA GeForce RTX 4060) is available, roughly how fast could this function run? Are there any benchmark results?
Does the execution time mainly depend on GPU performance, or are other factors (such as memory bandwidth or API overhead) more significant?

Test Environment

OS: Ubuntu 22.04.5 LTS (Jammy)
CPU: 13th Gen Intel Core i5-13400 (10 cores / 16 threads, up to 4.6GHz)
Integrated GPU: Intel UHD Graphics (PCI ID: 4682)
GPU Driver: i915 (Mesa, resolution 1920x1080)

[pioneer-innovation]

@fbxiang T_T I did not use ray tracing, just add a simple 640*480 camera on the end link of a 6-dof arm

[fbxiang]

get_picture has the following step.

1. Wait for the previous take_picture to finish rendering. (take_picture itself is async and returns immediately).
2. Copy the image from GPU to CPU.
3. Return the data as numpy arrays.

0.03s is a bit slow but considering it may involve rendering and transfer time, it is still reasonable. On NVIDIA GPU, we typically do not copy the data to the CPU but keep it on the GPU with get_picture_cuda so there is only rendering time and no transfer time. This typically results in 1-10 ms rendering time.

[fbxiang]

If you think there may be API overhead, it may help if you render multiple images all at once (many take_pictures from many cameras without get_picture). This can help the GPU to schedule the rendering more tightly with less API and transfer overhead.

[pioneer-innovation]

Thank you for the tip! At the moment I’ve only added a single camera.