Hi all,
I'm currently evaluating Beignet OpenCL performance on Skylake and Kaby
Lake GPUs against Nvidia CUDA, and surprisingly found that for our
specific task that the GPU performance when executing the same OpenCL
algorithms was substantially worse on a Kaby lake HD630 than it was on
a Skylake HD530. Is this to be expected?
The setup I am using is as follows:
- Ubuntu Xenial 16.04, kernel 4.8
- Beignet 1.3.0 compiled using the Zesty debian packaging on Xenial
- Intel i5-6500 (HD530) vs i7-7700 (HD630)
The OpenCL task is processing infrared data from Kinect 2 sensors,
using the Protonect tool from the libfreenect2 library two different
OpenCL-based algorithms: "default" and "kde" (the latter is more
computationally intensive).
Measuring GPU utilisation using the "render busy" metric from
intel_gpu_top, this is what I found:
HD530 utilisation:
- default: ~25%
- kde: ~52%
HD630 utilisation:
- default: ~35%
- kde: ~63%
I.e. the Kaby lake HD630 utilisation is 1.2-1.4 times heigher than the
Skylake HD530, however I would have expected the exact opposite.
Upgrading the kernel to 4.10 produced only a minor improvement for the
HD630. Would upgrading to Beignet 1.3.1 or a newer kernel than 4.10
help further at all? Any advice on how to substantially improve this
kaby lake performance would be appreciated.
Cheers,
//Mike
--
⊨ Michael Gratton, Percept Wrangler.
⚙ <http://mjog.vee.net/>
_______________________________________________
Beignet mailing list
Beignet@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/beignet
