Unable to use OpenCL on Linux

[Raestloz]

Fedora 31, Intel HD Graphics 520. I've updated to latest software updates, and I've installed Intel GPU OpenCL Runtime, as seen here:

Code: Select all

https://software.intel.com/en-us/articles/opencl-drivers#proc-graph-section

clinfo confirms I have OpenCL platforms, but for some reason FAHClient says clGetDeviceIds() returns -1. I've added fahclient user to video group and restarted, tried running FAHClient as sudo, but no avail

What can I do?

[bruce]

FAH does not support Intel Graphics devices. The OpenCL drivers from Intel are not useful.

FAH does support AMD and nVidia GPus and recommends you obtain OpenCL drivers from other sources.

[bruce]

While the CPU (Intel/AMD/etc.) can be configured as an OpenCL device, that's not useful either. FAH supports folding on the CPU without the need for OpenCL.

[Raestloz]

Hmmm, well CPU folding is a lot slower than GPGPU folding isn't it.

[bruce]

Generally speaking, that's true ... but the supported GPUs (from nVidia and AMD) are in a totally different class from the Intel Graphics devices. Intel designed them with a single purpose in mind: Displaying the images for a desktop application. Intel graphics is not designed to be a general purpose device to be used for the type of calculations that FAH and other Distributed Computing platforms need. Generally speaking, the intel graphics devices fill the narrow gap between the computational througput that a high-end CPU can provide and the low-end that can be provided by the oldest/slowest GPU devices available from nVidia/AMD.

That's a pretty narrow performance window -- hardly worth the support costs for FAH to develop a FAHCore that works on Intel Graphics devices.

If you want to fold on a GPU, my recommendation is that you add a graphics device that would be considered useful by gamers along with the OpenCL API that supports it.

[MeeLee]

Generally speaking, that's true ... but the supported GPUs (from nVidia and AMD) are in a totally different class from the Intel Graphics devices. Intel designed them with a single purpose in mind: Displaying the images for a desktop application. Intel graphics is not designed to be a general purpose device to be used for the type of calculations that FAH and other Distributed Computing platforms need. Generally speaking, the intel graphics devices fill the narrow gap between the computational througput that a high-end CPU can provide and the low-end that can be provided by the oldest/slowest GPU devices available from nVidia/AMD.

That's a pretty narrow performance window -- hardly worth the support costs for FAH to develop a FAHCore that works on Intel Graphics devices.

If you want to fold on a GPU, my recommendation is that you add a graphics device that would be considered useful by gamers along with the OpenCL API that supports it.

With Iris, that has changed.
Iris does 1Tflops (probably half precision, ~500Gflops full precision), has 32 to 64 cores running at 1+ Ghz.
Intel 10th gen CPUs have made a big difference.
But even their older gen (Boinc now accepts 8th and 9th gen CPUs), their GPUs more than double the work of the CPU.

I also think FAH has missed out big time, since most corporations and businesses, as well as many homes run Intel IGPs.
And even many of those that run AMD or Nvidia GPUs, still have an Intel IGP in their CPU.

[Joe_H]

It has not changed with Iris. On a performance level measured in Gflops the high end iGPU's from Intel are approaching parity with GT730 to GT750 cards. That is assuming the iGPU can keep up that level of performance for any length of time, they have to operate within the TDP for the entire CPU package.

Then there is the issue of drivers. Up to now the drivers and OpenCL support from Intel have not shown themselves as being up to supporting general GPU computational usage as needed for F@h. That is an important issue that would need to change for even minimal consideration of expending the time and effort to create a supported folding core to be used with the Intel iGPU's.

And yes, many homes and businesses run Intel powered systems with iGPU's. But most of those are not even to the level of the basic Iris, let alone the higher powered ones.

[bruce]

It has not changed with Iris. On a performance level measured in Gflops the high end iGPU's from Intel are approaching parity with GT730 to GT750 cards.

The GeForce GT Series cards, including the GT730 to GT750 cards are relatively low-end GPUs which can produce 700 to 1000 single precision GFLOPS 29 to 32 double precision GFLOPS. Most active GPUs are GTX series. (FAH doesn't use half precision.) These GPUs have been supported for many years and there is no particular reason to drop that support. Developing support for a new platform, however, isn't a trivial activity especially since it doesn't advance FAH's overall productivity significantly. I'm afraid you're right: FAH has missed the opportunity to develop support for the Intel Graphics platform.

More recent game-quality GPUs to produce upwards of 3000 to 10000 SP GFLOPS and 100 to 300+ DP GFLOPS and the recently released Radeon 5700XT produces 8000 SP / 500 DP.

Even Intel reports that many of the Iris iGPUs end up in systems with higher end add-on graphics cards.

[MeeLee]

There's still a market for lower end machines.
Boinc runs on Android, and it gets a significant contribution, despite phones being very weak in performance, even to IGPs...

I understand all the reasons of not implementing Intel in the race,
All I was saying it's just a shame they weren't able to do so.
Even 7th gen IGPs more than double CPU output, for just 10W more power consumption.
Heat is only an issue on laptops. Most PCs and mini PCs have enough cooling capabilities to keep temperatures at bay, even under full load.

[foldy]

For Intel iGPU gromacs CPU core latest version enables PME offload from CPU core to iGPU. So the gromacs CPU core uses the iGPU as accelerator. Maybe that is also possible for future FAHcore on CPU?

[bruce]

Maybe.

The fundamental question still is whether the Development team has the resources to create and test a new FAHCore and a new structure for servers that distribute specific projects to specific classes of hardware -- compared to the cost/benefit of other projects in FAH's backlog.

Clearly it's also a question of whether the development costs are as simple as replacing the latest version of GROMACS with the recent version used in FAHCore_a7 and deciding if the project being assigned needs different priorites compared to other projects based on the presence or absence of drivers for an Intel GPU -- (and whether FAHClient already reports that information to the server or if a new FAHClient would also have to be developed.)