Times (Elapsed/CPU) for BRP6-Beta-cuda55 compared to BRP6-cuda32 - Results and Discussion Thread

poppageek
poppageek
Joined: 13 Aug 10
Posts: 259
Credit: 2,473,733,122
RAC: 0

The times on my GTX 660 ti,

The times on my GTX 660 ti, Mint 17,has dropped from 12k seconds to 9-10k seconds.

http://einsteinathome.org/host/11667815/tasks&offset=0&show_names=1&state=3&appid=29

Cheers!

archae86
archae86
Joined: 6 Dec 05
Posts: 3,146
Credit: 7,100,804,931
RAC: 1,054,026

Here is my first result on

Here is my first result on the Windows CUDA55 application

A Windows7 GTX970 belonging to me has generated slightly over 22% productivity improvement based on the first six results.

As we observed on most of the Linux comparisons, this time again the CUDA55 timing distribution seems somewhat tighter than the CUDA32.

I admit the 6 results is a rather small comparison population, and in particular can't exclude a moderate high tail, so I intend to update this later.

Daniels_Parents
Daniels_Parents
Joined: 9 Feb 05
Posts: 101
Credit: 1,877,689,213
RAC: 0

First 8 BRP6 156 CUDA55

First 8 BRP6 156 CUDA55 (Win7, 2 GTX670) reported: Time saving 20 to 25 percent.

Arthur

I know I am a part of a story that starts long before I can remember and continues long beyond when anyone will remember me [Danny Hillis, Long Now]

Manuel Palacios
Manuel Palacios
Joined: 18 Jan 05
Posts: 40
Credit: 224,259,334
RAC: 0

I'm seeing a great

I'm seeing a great improvement in processing times with this application. I've gone from around 2:55h to 2:25h average runtime. I'll let more tasks run their course before I start doing any calculations.

ravenigma
ravenigma
Joined: 20 Aug 10
Posts: 69
Credit: 80,551,321
RAC: 13

Seeing about a 20%

Seeing about a 20% improvement in run time on a 980Ti. Slight (~2%) increase in power level.

Daniels_Parents
Daniels_Parents
Joined: 9 Feb 05
Posts: 101
Credit: 1,877,689,213
RAC: 0

First BRP6-Beta-CUDA5522

First BRP6-Beta-CUDA5522 Version 1.57 (Windows 32 Bit) successfully reported.

Arthur

I know I am a part of a story that starts long before I can remember and continues long beyond when anyone will remember me [Danny Hillis, Long Now]

archae86
archae86
Joined: 6 Dec 05
Posts: 3,146
Credit: 7,100,804,931
RAC: 1,054,026

with about a day and a half

with about a day and a half of CUDA55 work, the general appearance of my GTX970 comparison has not changed very much form the six sample graph I posted over a day ago.

This GPU runs 3X. The memory clock is strongly overclocked compared to stock, and the GPU clock is moderately overclocked. As both were running not far from their failure point on CUDA32, I am pleasantly surprised not yet to have seen a need to back down any for CUDA55 work, which one might have expected to be more challenging.

On this somewhat larger population than the one I first posted for this host, the calculated productivity improvement is over 24%.

Some tail populations have appeared, and it is not clear the time sample is comparable enough in whatever system variations drive the tails for a perfect comparison, but there is no doubt of a really big improvement, not much doubt that the primary distribution variability is less, and a least a hope that somehow the high tail has gotten less prominent.

archae86
archae86
Joined: 6 Dec 05
Posts: 3,146
Credit: 7,100,804,931
RAC: 1,054,026

My other two hosts each have

My other two hosts each have two dis-similar GPUs. As this dual arrangement robs performance available on a single GPU system, these results won't be directly comparable to the same GPU on single-GPU hosts. Both of the hosts include identical model GTX660s, but the hosts differ materially in motherboard, CPU, and non-BOINC usage. I'll present the two 660s in this post, allowing comparison, and the other two GPUs in another post.

The older host is actually my daily driver--on which I am typing this note. The GTX 660 card on this machine is scored with slightly over 24% productivity improvement from CUDA55. The reduction of high tail is noteworthy. It shares host resources with a GTX 750 and runs 2X.

The newer host sees negligible interactive use--it is primarily host to my backup drive and an alternate available to either me or my wife if our primary machine goes out of service. The GTX 660 card on this machine is scored with a 22% productivity improvement from CUDA55. The reduction of high tail is noteworthy. It shares host resources with a GTX 750 Ti and runs 2X.

archae86
archae86
Joined: 6 Dec 05
Posts: 3,146
Credit: 7,100,804,931
RAC: 1,054,026

RE: This post presents the

Quote:

This post presents the GTX 750 GPUs on the same two hosts running the GTX 660 GPUs in my immediate previous post. They are two different models of the same basic type of GPU, one a bare-bones 750, the other a 750 Ti.

The older host is actually my daily driver--on which I am typing this note. The GTX 750 card on this machine is scored with almost 33% productivity improvement from CUDA55. The reduction of high tail is very strong, and accounts for an appreciable share of the calculated improvement. It shares host resources with a GTX 660 and runs 2X.

The newer host sees negligible interactive use--it is primarily host to my backup drive and an alternate available to either me or my wife if our primary machine goes out of service. The GTX 750 Ti card on this machine is scored with a 17% productivity improvement from CUDA55. The reduction of high tail is apparent, but less than on my other observations. It shares host resources with a GTX 660 and runs 2X.


Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5,852
Credit: 111,160,961,088
RAC: 34,934,995

RE: ... In my last jobs

Quote:

... In my last jobs before retirement, I spent a great deal of time comparing data sets, and in cases like this am extremely fond of the probplot as a data representation that lets one spot many of the things one needs to see in order to avoid being misled by simplistic comparison.

....

For those (maybe all of you) unfamiliar with probplots, the data shown are the individual completion times in seconds, sorted in ascending order, with the actual time shown on the X axis, and the Z-score on the Y axis. A change of 1 in Z is one standard deviation, so the 0 Z value is the median. A normally distributed population will make a simple diagonal line from lower left to upper right.


Peter,

I really enjoyed your efforts to find specific hosts and publish the data for the types of GPUs that interested you. Thanks very much for doing that. I don't have such GPUs or I would have gladly published the data. I considered buying such a GPU but decided to hold off to see when/if NVIDIA will be releasing GTX 950/950Ti versions. At the moment, I'm very happy with my current fleet of AMD 7850 GPUs but there will be a day when I want to start replacing my NVIDIA 550Tis and 650/650Tis.

I must confess extreme ignorance about 'probplots', in fact the same is probably true for my understanding of anything but the most rudimentary levels of statistics in general. I did study the stuff in High School/University, but that was over 50 years ago and the stuff I learned then didn't really didn't get used at all in my working life - the old adage of "use it or lose it" certainly applies to me.

However Dr Google soon rectifies all ailments (I know I'll get a serve from a certain other Dr around here for that one :-)) and I was lucky to find an excellent primer to brush away the cobwebs. I found it very useful for that and the amazing thing is that it seems to have been written by a very dedicated person for the education system in Micronesia. It was particularly relevant for me because of the copious examples using LibreOffice Calc - which is exactly what I'm using, having never used MS Office products previously.

Apart from the linked document, I did quite a bit of browsing about normal probability plots - I assume you are wanting to test the data against a normal distribution rather than any other particular type of distribution - and I was interested to see that you can create such plots for other distributions as well.

My (very basic) understanding is that many things tend to be normally distributed and (apart from testing for the straight line probplot) another way of visualising the distribution is to plot a histogram of each data set. This looked like an interesting exercise so I've been accumulating data and working out how to organise things in LibreOffice to create a series of different views of the performance changes I'm seeing, particularly in Kepler series GPUs. I felt it was best to accumulate a reasonably large sample size and so this has taken a while. I'll try to start publishing some of this in the next day or so.

Cheers,
Gary.

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.