Now that the RTX 2070 is publicly available today along with already released 2080(ti) - is GPUGRID team working towards (Turing / Volta) CUDA10 support for near future?


FYI: RTX Turing Win8.1 isn't officially listed in Nvidia driver section. It will work if choose a Pascal card. Other then that fan support not supported on current Nvidia inspector but overclocking Turing is.

My RTX 2080 auto boosted to 2GHZ similar how Pascal was - where listed clocks are not always correct.

Now that the RTX 2070 is publicly available today along with already released 2080(ti) - is GPUGRID team working towards (Turing / Volta) CUDA10 support for near future?


FYI: RTX Turing Win8.1 isn't officially listed in Nvidia driver section. It will work if choose a Pascal card. Other then that fan support not supported on current Nvidia inspector but overclocking Turing is.

My RTX 2080 auto boosted to 2GHZ similar how Pascal was - where listed clocks are not always correct.

I haven't heard much from the team regarding the new architecture.

They took a while to support Pascal... so there's that.

---

http://ambermd.org/GPUPerformance.php

Turing is fast! RTX 2080 single job throughput bests the GTX 1080ti consuming fewer watts.

Sisoftware (scientist) benchmark have the RTX 2070 on par with GTX 1080ti.

GPUGrid Team:

Could you please respond to this thread?

GPUGrid Team:

Could you please respond to this thread?

Pretty please :)

---

I'm confused by this request. Why do you need CUDA10? All you need is the recent 410 series driver to support the card.

The card can run the CUDA8 app fine so you don't need a new application.

Unless, there is some function the current app uses that is incompatible with the Turing architecture. Only way to find out is to run a task.

The Einstein crunchers with Turing have found that some tasks error out immediately and other task types crunch fine. Bug report submitted to both Nvidia and Einstein developers.

Are Turing 20xx GPUS working in this project? I thought the answer was no, and that an app update from the GPUGrid devs was required. Keith do you have such a GPU to test?

Are Turing 20xx GPUS working in this project? I thought the answer was no, and that an app update from the GPUGrid devs was required. Keith do you have such a GPU to test?

No - ACEMD needs upgrade to run properly.

See the RTX 2080ti thread: http://www.gpugrid.net/forum_thread.php?id=4819

Currently: SWAN: FATAL: cannot find image for module [.nonbonded.cu.] for device version 750

Yes, I didn't remember that other thread which had the failures in the app because it wasn't compiled with API 7.5.0.

I just had the same question in the Seti forums where someone thought they had to run the CUDA10 version of an app just because the 410 drivers now have CUDA10 capability.

CUDA is backward compatible to all previous levels up to a point. The device itself and the drivers determine whether it will run the latest CUDA.

And of course, the applications need to understand CUDA10 too if trying to use that capability.

I tried the Seti CUDA10 app when it came out and found it 5% slower than the older CUDA9.1 or CUDA9.2 apps. So I went back to CUDA92 with the latest CUDA10 capable 411.73 drivers. Runs fine.

No I don't have any Turing card yet. Kicking myself for not grabbing the last of the 1080Ti cards. Now they all are gone.

Not in any hurry now since only Seti supports Turing. Neither Einstein or GPUGrid support it properly yet and I run those projects along with Seti. Not sure about MilkyWay. I have only one host running that project since it requires a more recent BOINC client than I normally run.

I see. I bet most of us are just wanting Turing app support, instead of CUDA 10 support ... And I thank you for pointing out that they are separate.

GPUGrid devs: Please update the app! :)

I see. I bet most of us are just wanting Turing app support, instead of CUDA 10 support ... And I thank you for pointing out that they are separate.

GPUGrid devs: Please update the app! :)

Yes please update! I have 2080 and 2070 ready for testing.

For Turing and Volta to work with GPUGRID - CUDA 10 required.

Since I've been crunching here - each major GPU release required an updated CUDA 5.0 / 5.5 / 6.0 / 6.5 / 8.0 CUDA ACEMD app.

When Kelper was released CUDA 5.0 was needed - same for 6.0 GM107 Maxwell - 6.5 for GM200/204/206 and 8.0 for GP102/104/106/108.

Though GPUGRID waited a year to adopt Pascal support.

For whatever reason GPUGRID never supported backwards compatibility.

The Einstein crunchers with Turing have found that some tasks error out immediately and other task types crunch fine. Bug report submitted to both Nvidia and Einstein developers.

I am afraid sending a bug report to NVIDIA won't help much.
I remember having read somewhere here in the Forum, some time ago, that NVIDIA explicitely states that their hardware is not supposed to be used for what we all are doing.

I am using a GTX 1060 on a Windows 10 PC and a GTX 750 Ti on a Linux box. They all work in SETI@home,Einstein@home and GPUGRID.
Tullio