GPU SP to DP ratio (as reported by GPU-Z 2.8.0, Advanced, CUDA)
Quadro 2000 1:12
Quadro 4000 1:2
Quadro 5000 1:2
gtx480 1:2
gtx1050ti 1:32
gtx1060 3gb 1:32
gtx1070 1:32

(RX550? RX480? GPU-Z 2.9.0, OpenCl, DirectX, Vulkan display no such information)

[QUOTE=kriesel;490571]GPU SP to DP ratio (as reported by GPU-Z 2.8.0, Advanced, CUDA)
Quadro 2000 1:12
Quadro 4000 1:2
Quadro 5000 1:2
gtx480 1:2
gtx1050ti 1:32
gtx1060 3gb 1:32
gtx1070 1:32

(RX550? RX480? GPU-Z 2.9.0, OpenCl, DirectX, Vulkan display no such information)[/QUOTE]

Anything about gtx980?

[QUOTE=ET_;490589]Anything about gtx980?[/QUOTE]

AFAICT 1:32.

GTX460=1:12 CC 2.1

gpu spec data
 

[QUOTE=kladner;490607]GTX460=1:12 CC 2.1[/QUOTE]
I've created the beginning of a gpu-specific reference materials thread at [URL]http://www.mersenneforum.org/showthread.php?p=490612#post490612[/URL]

If anyone would like to have such data as gpu-specific max temp, sp:dp ratio, cc level, etc tabulated, please make contributions in the reference material discussion thread [URL]http://www.mersenneforum.org/showthread.php?t=23383[/URL]
and please include a url for the NVIDIA or AMD or other spec sheet where it was found, so I can confirm the submission, and perhaps also tabulate other parameters not submitted.

(Sorry to have sort of hijacked the gpuowl thread with my list of gpu sp:dp ratios.)

[QUOTE=tServo;490563]The Kepler GPU based boards have excellent FP64 performance and are quite reasonable.
This would include the old standby original Titan and Titan black ( appx 300-400 us dollars on Ebay, used ), the Titan Z, the Tesla K80 ( 900 dollars ), and some of the higher-end Quadros of that architecture such as the K6000 and K5000.
The Quadros cost about the same as theTesla K80.
Most of these boards have 1/2 the FP64 math units as the FP32.[/QUOTE]

OK I stand corrected. Those are 3 generations back though, right? (are they Fermi, or Kepler, which was followed by Maxwell, Pascal, and Volta). No problem, it's good those were powerful GPUs, and they became affordable in time.

In the current generations, the most affordable Nvidia GPU with good FP64 I see is Titan V, which is about $4K.

[QUOTE=SELROC;489531]The problem here is not different program versions but an experimental system with experimental kernel and firmware, it is absolutely unstable and upgrades may break the system. At this time it was necessary (about 6 months ago) because the Drivers required some features of the latest kernel. So, I am going to refrain from upgrading this system further, gpuowl included, until those exponents are fully computed.

I am working to build a stable production system, this requires that the drivers work with the stable version of the kernel, the firmware also needs work. We can only wait until a new stable Debian release that works with latest drivers is out.[/QUOTE]


At least I now have a system with reproducible errors ...
So I went ahead and upgraded gpuOwl to the latest version, recompiling with the new g++-8.

I really have appreciated the current backward-compatibility, as it restarted from the last savefile.

note: openowl -h doesn't show the list of devices.

[QUOTE=SELROC;490712]
note: openowl -h doesn't show the list of devices.[/QUOTE]


Yes, I'm aware of that, it's because of the recent backend split.

[QUOTE=preda;490739]Yes, I'm aware of that, it's because of the recent backend split.[/QUOTE]


But I think that openowl is becoming robust and fairly reliable.

[QUOTE=preda;490739]Yes, I'm aware of that, it's because of the recent backend split.[/QUOTE]




By now you should have to detect the two types AMD and NVIDIA ... I don't know if this requires different instructions ...

[QUOTE=SELROC;490762]By now you should have to detect the two types AMD and NVIDIA ... I don't know if this requires different instructions ...[/QUOTE]

Do you have a computer with both AMD and Nvidia GPUs?

That would require an executable that is linked with both OpenCL and CUDA. That can be done, I just wasn't sure if there's a need for such a build.