nvidia rtx-2080

[nomead]

Originally Posted by kriesel

Who installed your son's gpu? With careful antistatic precautions?
(Way to get our hopes up and then dash them, by the way...)

I installed everything... I work in electronics manufacturing so ESD precautions are pretty much second nature to me... at work. But at home I tend to wing it. Still, some basic procedures I adhere to.

But happy me, it wasn't the display card after all. The problem only manifested itself as a BSOD in pci.sys ... but the real reason was that memory clocks were set too high. Ah well, that's what you get when you're too eager to play and don't let me run a Prime95 torture test first...

[kriesel]

Originally Posted by nomead

I installed everything... I work in electronics manufacturing so ESD precautions are pretty much second nature to me... at work. But at home I tend to wing it. Still, some basic procedures I adhere to.

But happy me, it wasn't the display card after all. The problem only manifested itself as a BSOD in pci.sys ... but the real reason was that memory clocks were set too high. Ah well, that's what you get when you're too eager to play and don't let me run a Prime95 torture test first...

...or memtest86+ overnight. Congrats on sorting it out. That's good news it is not a bag gpu card.

[kriesel]

Originally Posted by nomead

Ok the path of least effort is probably as an attachment, then. So, here it is.

Any chance you could knock out a CUDA8 Windows build of the 2047-gpusievesize mod?
I could bump a system up from CUDA8-capable driver to CUDA10-capable, to use this on a gtx1080Ti, but that would deactivate the older gpu in the system. And similar to that is the case on more than one system here.

[chalsall]

Originally Posted by kriesel

Any chance you could knock out a CUDA8 Windows build of the 2047-gpusievesize mod?

Interesting. A call for an executable?

Is it as difficult to get a native build for Windows environments as it is for a Linux variant?

Sincere question.

This shouldn't be difficult. Empirically it can be,

[nomead]

Originally Posted by chalsall

Interesting. A call for an executable?

Is it as difficult to get a native build for Windows environments as it is for a Linux variant?

On Linux it isn't, just make. On Windows... it's a bit fiddly. I'd say it's much more difficult than on Linux. But I still have the environment up and running, so I'll try it at some point today.

[kriesel]

Originally Posted by chalsall

This shouldn't be difficult. Empirically it can be,

Last time I tried it, the link step complained that one of the many pieces was 32-bit while the rest were 64-bit, instead of succeeding, and WHICH piece was the exception was a well kept secret.

[nomead]

Okay, here goes... I actually had to read my own notes from half a year ago, reinstall the 8.0 version of the CUDA toolkit, and even then, only the third attempt successfully produced executables.

GPU_SIEVE_SIZE_MAX was set to 2047 in params.h
64-bit binary, compiled for both more and less classes. Code generated for these compute capabilities: 30 35 37 50 52 60 61 62 (Unfortunately with CUDA 8.0, anything older than CC3.0 and anything newer than CC6.2 isn't supported)

Also unfortunately I don't have any hardware to test this with... got nothing between GT430 and RTX2060

[kriesel]

Thanks nomead, this works, in the sense of appears to run properly, and also provides a considerable speedup on my gtx1080Ti. Based on benchmarking there, it looks like if the program was revised to support unsigned-32-bit GpuSieveSize up to 4095 Mibibits, there would be additional slight gains. Running more than one instance adds a small throughput increase.

[Xyzzy]

https://developer.nvidia.com/unrealengine

https://developer.nvidia.com/rtx-tech-showcase