Panorama
Nvidia CUDA-only
Linux and Windows
Nvidia driver version of 418.96 or higher.

Kaktwoos
AMD, Intel and Nvidia GPU
AMD and Intel require OpenCL 2.0
Nvidia requires OpenCL 1.2


AMD GPU Radeon GCN 1.2+
RX 400 series
RX 500 series
Radeon drivers 19.12.3

RX Vega series
Radeon VII
RX 5000 series
Possibly others like Radeon Pro and FirePro W series
Radeon drivers 20.5.1 or better may be required

Intel GPU
Intel Iris Plus Graphics (OpenCL 2.1)
Intel UHD Graphics 610 (OpenCL 2.1)
Intel HD Graphics 610 (OpenCL 2.1)
Intel UHD Graphics 610/620/630/P630 (OpenCL 2.1)
Intel HD Graphics 610/615/620/630 (OpenCL 2.1)
Intel Iris Plus Graphics 640/650/655 (OpenCL 2.1)
Intel HD Graphics 520/530 (OpenCL 2.0)
Intel Iris Plus Graphics 540/550/580 (OpenCL 2.0)
Intel HD Graphics 6000/5500 (OpenCL 2.0)
Intel Iris Graphics 6100/6200 (OpenCL 2.0)
Intel HD Graphics 5300 for Intel® Core™ M Processors 515/5300 (OpenCL 2.0)
These are only available on generation 5 (Broadwell), generation 6 (Skylake) and above Intel CPUs.


Nvidia
Nvidia is a bit of a mismatched clusterf*** when it comes to OpenCL support.
While initially they embraced the new open standard, they stopped support for newer versions because it is a direct competitor to their own CUDA.

Basically, the earliest models like a GTX 465. 470 and 480 have OpenCL 1.2 support, while all of the GT(X)500s do not.
All of the GT(X)600s, most 700s, all 900s, all 1000s and all RTX are OpenCL 1.2 compliant.

I'll leave this thread open so people can add to it, or ask questions.

Linux drivers:
HY, MAH developer wrote:

For Intel:
Ubuntu 18.04/18.10: (intel) beginet beginet-opencl-icd (works for my UHD 600 on 18.10)
Ubuntu 19.04/20.04: intel-opencl-icd ("NEO")

For AMD:
ROCM for 18.04 and newer for Open-Source drivers
https://rocmdocs.amd.com/en/latest/Installation_Guide/Installation-Guide.html

Or, proprietary AMD drivers can be used if they are already installed. I personally have no experience with the proprietary Linux drivers
https://linuxconfig.org/install-opencl-for-the-amdgpu-open-source-drivers-on-debian-and-ubuntu

For Nvidia:
Nvidia official/proprietary drivers of 400 or newer, but for Panorama (cuda) 418 is required currently. After installing the Nvidia drivers on Ubuntu, you *must* install the nvidia-opencl-dev packages or the related OpenCL Nvidia headers on your distribution. This has caused major errors due to the OpenCL driver being installed by the package manager, but *not* what helps kaktwoos actually interact with it

It ran successfully earlier in July on Intel HD4600:

https://minecraftathome.com/minecrafthome/result.php?resultid=2711899
https://minecraftathome.com/minecrafthome/result.php?resultid=2718844

|  | Processor: 8 GenuineIntel Intel(R) Core(TM) i7-4720HQ CPU @ 2.60GHz [Family 6 Model 60 Stepping 3]
...
|  | OpenCL: Intel GPU 0: Intel(R) HD Graphics 4600 (driver version 20.19.15.5126, device version OpenCL 1.2, 1630MB, 1630MB available, 192 GFLOPS peak)

Paired with a GTX 960.

Got these right after the fix for the WUs not obeying the GPU run prefs on the client. Haven't got anymore for Intel since then though.

So, the truth is it actually never worked

But, because you had an Nvidia GPU, a bug with our older multi-gpu on Kaktwoos_2.00 meant all Intel tasks ran on the first Nvidia or AMD GPU.
If people with only an Intel 4000/4600 ran the Kaktwoos_200 or newer before we enforced OpenCL 2.0, it'd fail due to its older OpenCL implementation and missing features that we do use. Saw lots of errors and hence the specific requirements posted above (thanks, Jord)


A small thing to keep in mind for AMD, we cannot support all 200/300 Radeons because some of those models have Pitcarin/Tahti GCN 1.0/1.1 silicon, and not GCN 1.2+ (Hawaii/Bonair)

I recommend for Vega, Polaris (RX 470 -> RX 590) and older that the AMD 19.12.3 drivers or slightly older than that are used (if on Windows). For RX 5000 Navi and Radeon VII, we recommend 20.5.1, though some users can test 20.7.1 if they wish (some bugs reported)

Driver support is complicated though for Linux
For Intel:
Ubuntu 18.04/18.10: (intel) beginet beginet-opencl-icd (works for my UHD 600 on 18.10)
Ubuntu 19.04/20.04: intel-opencl-icd ("NEO")

For AMD:
ROCM for 18.04 and newer for Open-Source drivers
https://rocmdocs.amd.com/en/latest/Installation_Guide/Installation-Guide.html

Or, proprietary AMD drivers can be used if they are already installed. I personally have no experience with the proprietary Linux drivers
https://linuxconfig.org/install-opencl-for-the-amdgpu-open-source-drivers-on-debian-and-ubuntu

For Nvidia:
Nvidia official/proprietary drivers of 400 or newer, but for Panorama (cuda) 418 is required currently. After installing the Nvidia drivers on Ubuntu, you *must* install the nvidia-opencl-dev packages or the related OpenCL Nvidia headers on your distribution. This has caused major errors due to the OpenCL driver being installed by the package manager, but *not* what helps kaktwoos actually interact with it

Now, even if you don't have a compatible graphics card:

For those with a multi-core, relatively modern (Bulldozer/Phenom and Intel Nehalem) and newer CPU, a version of kaktwoos is being made to run on CPU only systems or those with incompatible GPUs. It will serve the same purpose as kaktwoos-cl that all of us are running atm, but it's actually a decently fast alternative.

On my Ryzen 1700 (8c/16t, 3.75GHz all cores) I get a speed of ~20mseed/s if nothing is running and I set it to use all of the processor. This is comparable to a RX 580 on our latest kaktwoos-cl version!

A Intel 8600K runs at 12-13mseed/s, and I expect a Ryzen 2600 to perform similarly

On a Ryzen 3950x or older 16 core threadripper, we can expect 40 million seeds a second, at 150W; comparable to my desktop's overclocked Vega 56.

Sadly, a Raspberry Pi 4 at 1.95Ghz, 4 cores (32bit OS) runs this at... 1.2mseed/s

If you would like, progress of our CPU kaktwoos application can be tracked here:
https://github.com/DutChen18/kaktwoos-rs (Orig)
https://github.com/Hyenadae/kaktwoos-rs/commits/master (BOINC native mods + auto build for linux and soon Windows)

Just curious if it ran on the nVidia chip, why did it have a notably longer time crunching?

I didn't try running both at the same time. I ran one chip app than the other so I didn't get to see if they actually both tried to run at the same time.

I'm just wondering if it was intended for the Intel GPU how it got switched over to the nVidia GPU? I know Folding@home that kind of weirdness happens, but I thought BOINC (short of manually hacking your settings) didn't do that?


Also, the Intel/AMD OpenCL 1.2 support, is that a hardware issue or a driver issue?

Also, the Intel/AMD OpenCL 1.2 support, is that a hardware issue or a driver issue?

OpenCL is hardware. You cannot update it with drivers on most systems. Some AMD FirePro can be updated from 1.2 to 2.0 under Linux with specific drivers, but that's because the hardware was already 2.0 capable.

Might be thinking OpenGL... Can't we just pick one name already? Guess you can't trademark OpenHouse. ;)

But for that OpenGL is also hardware. Even DirectX is hardware. Though with these it's usually the software that is the limiting factor. For instance you require Windows 10 for DirectX 12 compliancy, any Windows before that is DirectX 11 or lower. What the drivers do is tell the operating system what the hardware can do, but they can't tell the operating system things the hardware cannot do. So while you can install OpenCL 3.0 drivers or OpenGL 4.6 drivers or DirectX 12 Ultimate drivers, when you hardware doesn't have the capability to run those it will default to the version the hardware can do.

OpenGL may be deprecated in favour of Vulkan, but that may not happen on all platforms.

I think around DirectX 5 that going forward to newer DirectX would sometimes cause weird things to happen with games. It's been eons since I played games though, so I don't know if that's improved any.

I thought I remember reading about either (or both) OpenGL / OpenCL not being properly supported by the drivers which is why I was unsure.

These days, if stuff doesn't run I sigh (usually blaming Microcrash err... soft) and going looking online to see why it doesn't and hopefully finding a solution. Folding@home seems more sensitive to nVidia driver updates than BOINC projects. Just something in the way Folding@home implements GPU computing.

I'm still curious if it could be explained in layman's terms what Intel and AMD are lacking in supporting OpenCL 1.2 properly that you need to use the newer version. Is this happening on other projects too? I don't have an AMD, and there's a lack of Intel GPU science projects so I haven't done any Intel GPU in a whiles (especially with the nVidia being preferable).

Cheers,

Yav

Not sure if anyone will see this with a new post, but we've been doing some calculations today for some things we are working on atm:

Per-GPU "TFLOPS vs Seeds" architecture comparison

15,276.52 GFLOPS / 874.033 seconds (Turing)
2,178.28 GFLOPS / 6104.476 seconds (Pascal)
6,313.96 GFLOPS / 5216.085 seconds (Polaris)
11,401.80 GFLOPS / 2586.161 seconds (Vega)
9,500.00 GFLOPS / 1136.033 seconds (Navi)

15,276,520,000,000 flops / 114,410,000 = 133524 FLOPS per seed on Kaktwoos-cl (Turing RTX 2080ti )

2,178,000,000,000 flops / 16,380,000 = 132967 FLOPS per seed on Kaktwoos-cl (Pascal GTX 1050ti )

6,313,960,000,000 flops / 19,170,000 = 329367 FLOPS per seed on Kaktwoos-cl (Polaris RX 580)

11,401,800,000,000 flops / 38,670,000 = 294849 FLOPS per seed on Kaktwoos-cl (Vega 56 OC)

9,500,000,000,000 flops / 88,030,000 = 107918 FLOPS per seed on Kaktwoos-cl (Navi RX5700XT)

TLDR, with current Kaktwoos-cl optimizations, Turing and Navi GPUs are by far the best architectures "Per TFLOP" to calculate seeds. Obviously, this data is of a small sample size and also, RX 5700XT TFLOPs are incorrectly reported by BOINC as "4.5" rather than the reported 9.5tflop by the device specs. May be because of the RDNA architecture splitting compute units in 'half' versus Polaris and Vega?

We also can now make a better estimate on average for how much real compute power is needed per Kaktwoos-cl task as well

Do we have instructions for Kubuntu/Linux Intel GPU users yet? Beignet OpenCL stuff doesn't seem to do the trick for you all.

---

https://stats.free-dc.org/cpidtagb.php?cpid=41442452452b2202db624a3666783bab&theme=10&cols=1

Also, the Intel/AMD OpenCL 1.2 support, is that a hardware issue or a driver issue?

OpenCL is hardware. You cannot update it with drivers on most systems. Some AMD FirePro can be updated from 1.2 to 2.0 under Linux with specific drivers, but that's because the hardware was already 2.0 capable.

My R9 280x, under Windows 10, had OpenCL 2.0 compliance under an older driver. They latest driver forced it back to 1.2.

I'll have to rediscover which of the downloaded, multitude of drivers versions 18.x.x -> 19.x.x gave me the OpenCL 2.0 capability or if I needed to do something special to get it turned on.

I know Linux users had to resort to special drivers to get 2.0 for that series but some of the Windows versions had it.