Projects 762x Testing Core v2.25 on Adv

[tjlane]

Folding Forum,

Please be aware that we are rolling out a new version of GPU Core15, v2.25. This core includes support for the new generation of nVidia boards, the Kepler series.

While this core has been extensively beta-tested and not shown issues, we are doing a controlled release to the general FAH community. Over the next few days, the p762x series will run on advmethods and require core15 v2.25. Your client should upgrade automatically to the new core.

Please let us know by posting here what your experience with the new core is -- if all goes well, we will release to the entire FAH community shortly, so we can support Kepler boards.

Thanks!

TJ

[7im]

Supported does not equal optimized.

Any comments in regards to speed? Or further optimizations possible?

[k1wi]

I haven't noticed any increase in performance from v2.25 on my 500 series GPU, are you referring to Kepler 7im?

[7im]

I haven't noticed any increase in performance from v2.25 on my 500 series GPU, are you referring to Kepler 7im?

Yes.

... This core includes support for the new generation of nVidia boards, the Kepler series.

Also concerned that v2.25 caused slightly longer frame times on a few projects using Fermi boards.

[Zagen30]

The core was updated without issue. I'm currently running 7626 (Run 103, Clone 0, Gen 83) on one GTX 460 and 7624 (Run 111, Clone 0, Gen 88) on the other. HFM reports ~9:13 TPF resulting in ~8k PPD for both. My previous history (a few hundred WUs according to HFM) with 762x was that they ran at about 6:50 TPF for ~11k PPD.

[Grandpa_01]

I haven't noticed any increase in performance from v2.25 on my 500 series GPU, are you referring to Kepler 7im?

Yes.

... This core includes support for the new generation of nVidia boards, the Kepler series.

Also concerned that v2.25 caused slightly longer frame times on a few projects using Fermi boards.

Actually the core causes much longer frame times on some projects, On my GTX 580 Project 8018 core 2.22 = 2:51 core 2.25 = 4:59 Hopefully this will not cause ther rest of the folding comnity problems just to support Kepler.

[k1wi]

The core was updated without issue. I'm currently running 7626 (Run 103, Clone 0, Gen 83) on one GTX 460 and 7624 (Run 111, Clone 0, Gen 88) on the other. HFM reports ~9:13 TPF resulting in ~8k PPD for both. My previous history (822 WUs according to HFM) with 762x was that they ran at about 6:50 TPF for ~11k PPD.

Ah damnit - now that I've taken a closer look I get the same drop as grandpa on project 8018 on my 570. There goes my blissful unawaredness!

The lower PPD is similar to the PPD I got with 2.22 on the 76xx projects, so I wonder if that means their ppd is even lower again with 2.25 or whether there is some sort of interaction going on.

[Grandpa_01]

yes the 762x series will have the same reduced tpf as the 8018's and all the other WU's with 2.25

[rhavern]

I concur with the increased TPF, checked HFM, on the same hardware GTX460 OC to 1600 shaders, P8018 2.22 is 5:31 whereas 2.25 is 7:11. That is non-trivial.

[BigApplePi]

Pardon me but I'm just a novice with this. I had been running a 680, model 02G-P4-2682 successfully for a few months now. Suddenly at the time of your current announcement, the gpu project given 7623 doesn't work. It says it's running but progress stays at zero. A cold reboot of the computer doesn't help. No automatic start. Advise please. I removed the gpu WU but don't know how to add it back again.

Upgraded version 7 to 7.2.9, found the instructions for removing and adding slot. Got new WU and am okay now.

[Napoleon]

This is the exact same core v2.25 which was discussed in beta forum: viewtopic.php?f=66&t=21571, right? The tpf increase was also noted there. Furthermore, there's this statement by friedrim in viewtopic.php?f=66&t=22197&p=221635#p221635:

In the past there were two Cuda Cores based on OpenMM: Core_15 and Core_16. The Core_15 had an optimization that is no longer being applied, and Core_15 projects are being phased out. The optimization associated w/ the Cuda Core_15 was never implemented on the OpenCL platform so there has only been a Core_16 for OpenCL.

With core v2.22 my GT430 gets about 3500 PPD for P762x WUs. With core v2.25 it only gets about 2500 PPD. So yes, looks like some optimization was dropped from v2.25. P762x WUs + core v2.25 is a tough combination for my GT430. P762x WUs themselves produce about 1000 PPD less compared to other projects, and the new core shaves off yet another 1000 PPD. At the moment I'm in the middle of a P8018 WU / core v2.22, with an estimated 4800 PPD.

Core_15 v2.25 is the new FAH core for Fermi & Kepler GPUs. This is actually a bit confusing, at least to me. "Core_15 projects are being phased out"

[Napoleon]

Any comments in regards to speed? Or further optimizations possible?

I noticed that the new 310.33 beta driver (released Oct 23, 2012) supports the CUDA 5 Toolkit (released Oct 15, 2012). I feel quite awkward, yammering about CUDA 5 in this very thread, but if I've understood e.g this article correctly, speedups will come when some FahCore after v2.25 adopts CUDA 5.

Developer Accolades for CUDA 5
Developers who evaluated the pre-release version of CUDA 5 have reported often dramatic application acceleration and improved programmability.

[7im]

That's what people have been trying to say all along. PG is waiting on OTHER developments to optimize NV and ATI speeds.

[Napoleon]

Indeed. Which is why I pointed out - more or less subtly - that first CUDA 5 capable driver was released only a couple of days ago, and a beta driver at that. Not to mention that the core also needs to be adapted to make most of the new CUDA features.

[bruce]

FAH Development may or may not be running internal tests with a beta version of the drivers but they DO NOT recommend that Donors use beta versions. (It's possible that beta drivers might create invalid results which might then contaminate the store of valid scientific data.) Certainly Development is not going to distribute a FahCore (even to beta testing) unless it works with drivers that have been formally released by Nvidia/AMD.