AMDGPU-PRO 17.10 OpenCL vs. NVIDIA Shows Problems

[wizard69]

Hopefully Ryzen wil turn things around at AMD and they will have the cssh to hire a few programmers. In the end this looks like a software problem not a hardware problem. Frankly AMD needs to invest in problem areas like this to regain credibility.

Possibly more important they need to acknowledge the issues.

[juno]

This is why I recently bought a NVIDIA GTX 1060. What is Open Source good for, when both, the open source driver and the blob driver from AMD just suck?!

Where does the free stack suck?
Yeah, it lacks features but what works does work way better than anywhere else.
You want to know what sucks? Nvidia's ancient intrusive blob stuff does. Oh, and it's lacking features, too.

OK, your driver sucks for everyday tasks except for a few games. But hey, you get to use GeForce Experience with forced registration, inbuilt telemetry. And you get a running node.js server that tears a hole into your system with your damn graphics driver stack. Now that's great! Today on Windows - soon on Linux.

[flubba86]

These results reflect my experience using OpenCL accelerated Cycles in Blender with the AMDGPU-PRO driver. With all settings tweaked for optimal performance, it is barely faster than CPU rendering, and much slower than equivalent nVidia cards. And CUDA backend (only on nVidia cards) is still way faster, but that is a discussion for another thread.

[smitty3268]

AMD's OpenCL and Vulkan support is a disaster right now. At least their GL driver is finally rounding into shape.

[patstew]

Don't worry, AMD have scheduled working drivers to be ready in time for the year of the linux desktop

[nevion]

AMD's OpenCL and Vulkan support is a disaster right now. At least their GL driver is finally rounding into shape.

OpenCL via ROCM is coming up, stretch to say disaster as it's just on the cusp of being awesome. 1.5 coming out later this week and I'm eager to look into it.

[eydee]

Don't worry, AMD have scheduled working drivers to be ready in time for the year of the linux desktop

I heard they'll be released the same day as HL3.

[Steffo]

Where does the free stack suck?
Yeah, it lacks features but what works does work way better than anywhere else.
You want to know what sucks? Nvidia's ancient intrusive blob stuff does. Oh, and it's lacking features, too.

OK, your driver sucks for everyday tasks except for a few games. But hey, you get to use GeForce Experience with forced registration, inbuilt telemetry. And you get a running node.js server that tears a hole into your system with your damn graphics driver stack. Now that's great! Today on Windows - soon on Linux.

I don't know what you are talking about. Forced registration? What?! node.js server?! I really don't know what you are talking about!
And why should I buy a dedicated graphic cards when the performance and the features suck?
I do not only play games sometimes, I use the graphic cards for scientific computations, too. So, why should I buy AMD when they suck on both areas?!

[jlavi]

Probably no one uses it on unsupported distro, since it is broken .

I am using MSI RX 480 8GB, v4.10 kernel and Devuan and using the card to accelerate 3D modeling.

In 17.10 I edited the regular expression on one line in amdgpu-pro-install to accept also my distribution. After that I ran "sudo bash amdgpu-pro-install --compute -y". DKMS package installs but won't compile against v4.10 kernel. Yet the OpenCL works but performance fluctuates.

I ran the Shoc Triad test in a busy loop and the results varied in the range 4.0 .. 10.7 GB/s so that most (about 19 per 20) were in the low end and every now and then some value was in the high end. When running the Triad tests in a busy loop the card temperature stayed at 40C with fans spinning. The idle temperature for the card is about 33C.

In the article they got 8.39 GB/s for RX 480 in SHOC Triad test.

Something (internal clock? internal synchronization?) slows down or speeds up the results at random. The benchmark should be made aware of this issue.

In Shoc FFT SP test I got 527 .. 535 Gflops/s tested in a loop. In the article they got 4.05 Gflops/s and in January 508 Gflops/s with 16.60 driver (article: NVIDIA Linux OpenCL Performance vs. Radeon ROCm / AMDGPU-PRO).

[iszotic]

These results reflect my experience using OpenCL accelerated Cycles in Blender with the AMDGPU-PRO driver. With all settings tweaked for optimal performance, it is barely faster than CPU rendering, and much slower than equivalent nVidia cards. And CUDA backend (only on nVidia cards) is still way faster, but that is a discussion for another thread.

Cycles is working fine with this driver, I'm getting the benchmark results on the next 2.79 release with 16.04 LTS and kernel 4.4.0-75, I rendered Koro scene in 540secs. The cl-mem tests are working and are around the gtx 1070.