Another impressive factor .... I just had a driver reset, my fault I was messing with a few things. It all paused as the driver reset itself, and carried on like nothing happened - all WUs running at time of reset validated, and all back to normal without me doing a thing.

If thats not a one off fluke ... thats brilliant, no more getting up in the morning to find never ending WUs cycling for nothing for hours on end in the night as a result of a reset in the night.

Regards
Zy

BTW, the page below needs to be corrected and updated:

http://milkyway.cs.rpi.edu/milkyway/gpu_list.php

The HD 69xx series isn't even listed (unless the top AMD listing of HD5800 is a typo) and no the GTX480 is not faster than the GTX580 according to the WU completion times in "Top Computers".

I'm guessing that since the ATI branding was dropped for 6xxx and the BOINC script seems to look for things with "ATI" in the name it never finds it.

It's interesting that even at this point the HD 7970 is running the NVidia optimized app at 3.5x faster than the top GTX580 card that I could find in the database. Pretty impressive and the HD 7970 also uses less power.....

Also interesting is the low CPU usage. Some other OpenCL GPU applications use a lot more CPU.

Run Time is 52 secs with GPU set to 1125/1575, the max speed inside the driver, power slider is +5

I hope this is only the beginning without optimized applications. This is small better than the performance of a 6970 at stock, look here: http://milkyway.cs.rpi.edu/milkyway/results.php?hostid=357618&offset=0&show_names=0&state=3&appid= (Thanks to joe!)

It's interesting that even at this point the HD 7970 is running the NVidia optimized app at 3.5x faster than the top GTX580 card that I could find in the database. Pretty impressive and the HD 7970 also uses less power.

I´m not surprised. The customer versions of Fermi are crippled on the DP-performance. Good for this project! I´m interested in the comparison on SP-projects like collatz or primegrid or maybe in a long, long, long future GPUGrid.

---

Would be interesting to see how the NVidia OpenCL app runs on a 6970 .... that would give an early level playing field comparison. I suspect it would be very significantly slower.

Regards
Zy

I ran the cards overnight @1115/300 to see if low memory worked (I run my 5970s @770/175). The invalids came in, whilst still a very small number, they were still there.

At 1030am I changed it to 1125/1000 (default is 1125/1375), and the invalids have pretty well stopped - insignificant now. So its not a happy bunny down at 300, probably not surprising given the 4.3Bn transistor count.

So there is a sweet spot between 300 & 1000 on memory for MW on this NVidia app, dont yet know where, not worth fiddling more for the moment, going to just let it run @1125/1000. The latter is a safe start point for anyone getting a 7970.

For sure, staying down at 925 gpu is a travesty on this card, wack it up to 1125 immediately, and power slider to 20%. Then play with memory as needed. The power slider only raises the maximum TDP, it does not mean the extra power is always on just makes it available, so may as well wack it to the +20 and leave it there, that eliminates low power levels as a point of failure when dealing inside the driver settings.

Regards
Zy

I ran the cards overnight @1115/300 to see if low memory worked (I run my 5970s @770/175). The invalids came in, whilst still a very small number, they were still there.

At 1030am I changed it to 1125/1000 (default is 1125/1375), and the invalids have pretty well stopped - insignificant now. So its not a happy bunny down at 300, probably not surprising given the 4.3Bn transistor count. Zy

Sometimes it takes a while for the new invalids to show up. All my machines are getting a 2-3% invalid rate. When I start MW on a new box there are no invalids for a while, then they start dribbling in. They're all of the sderr bug type and I think a lot of them get held in "pending" until they're finally marked as invalid. Even Matt's machines show the same problem and it's one of the 3 reasons I dropped MW for a while (along with the tiny queue and unreliable servers). Since the queue size is up I'm back to give it another try but the false invalids are unique to MW and quite irritating.

sesef managed to finish MW WU in 35-36 sec. we are testing now.

with these drivers http://forums.guru3d.com/showthread.php?t=356972 i have significant speedup, but there is bad news :( After drivers update I can't control card with Afterburner :(

Try Sapphire Trixx - works on reference AMD cards

https://www.sapphireselectclub.com/ssc/TriXX/TriXX.aspx

Regards
Zy

why not use the official driver from AMD's site dated 1/9/2012?
http://support.amd.com/us/kbarticles/Pages/radeon-7900-series.aspx

I'm trying to run the cuda_opencl app on my 6970 just for comparison.

I used the app_info described here and downloaded the app from the link posted, but boinc says

1/14/2012 12:19:06 PM | Milkyway@home | Sending scheduler request: To fetch work.
1/14/2012 12:19:06 PM | Milkyway@home | Requesting new tasks for ATI
1/14/2012 12:19:08 PM | Milkyway@home | Scheduler request failed: HTTP file not found

I'm trying to run the cuda_opencl app on my 6970 just for comparison.

I used the app_info described here and downloaded the app from the link posted, but boinc says

1/14/2012 12:19:06 PM | Milkyway@home | Sending scheduler request: To fetch work.
1/14/2012 12:19:06 PM | Milkyway@home | Requesting new tasks for ATI
1/14/2012 12:19:08 PM | Milkyway@home | Scheduler request failed: HTTP file not found

You can try it zydor wrote in another thread.
A detatch (not reset) and reattatch will do it.

I'm trying to run the cuda_opencl app on my 6970 just for comparison.

I used the app_info described here and downloaded the app from the link posted, but boinc says

1/14/2012 12:19:06 PM | Milkyway@home | Sending scheduler request: To fetch work.
1/14/2012 12:19:06 PM | Milkyway@home | Requesting new tasks for ATI
1/14/2012 12:19:08 PM | Milkyway@home | Scheduler request failed: HTTP file not found

You can try it zydor wrote in another thread.
A detatch (not reset) and reattatch will do it.

thanks

Renaming the app into _ati14 also did the trick (done also when detaching, so i'm not sure what benefitted more)

Anyway, this is my result: 1m 20s - 1m 25s, while also doing some cpu computation. CPU usage is 0, screen is a bit lagging.

Link to my host:
http://milkyway.cs.rpi.edu/milkyway/results.php?hostid=364107

On the contrary, I resumed working on milkyway for some wus and the times with stock (CAL) app are around 45-55 seconds, with 0 cpu usage and the screen absolutely better than when using the opencl app.

Let's hope that some parameter tuning within the opencl app could bring similar results to the cal app. Surely it will not be better than that, but at least maybe closer... Let's hope, also because the cal interface is deprecated and so this old app will soon stop working with newer drivers...

let's hope that some parameter tuning within the opencl app could bring similar results to the cal app

Sadly, in pure terms of OpenCL, that will not happen to the level past experiences may lead you to believe is possible, because OpenCL code is compiled by Vendor/OpenCL compilers. That means little detailed tweeking - if any - is possible such as diving in with assembler coding, as it is all brought back to a level playing field by the compiler. The latter makes sense when looking at what OpenCL is for - cross platform compatibility - but in the short term presents some issues in niche areas such as BOINC. Fundamental speed improvements will be (as such given tight coding in the first place) dependent on the effort and development of the compiler by the Vendors ......

You saw the difference in speed on the 6970 run compared to CAL, and the fact that the 7970 is doing it in 49-52 seconds, compared to 90 seconds+ with the 6970. That amply demonstrates the hardware leap forward by 7XXX as well as starkly showing the transition issues in terms of speed CAL/OpenCL.

As to the future .... well the new GCN architecture is not yet even started being properly brought into play, and it (realistically) promises much. It will however take time for developers to wrap their heads around the new Architecture and code around the whole new mix.

Its likely therefore that new BOINC AMD apps will take a speed hit as they will be OpenCL. I wouldnt rush off and knee jerk an NVidia buy just on the back of that however, as they will also face similar issues as the industry is moving to OpenCL as the "standard". Will there be "cheating" and "exceptions", of course there will be - its the nature of the beast - but such short term grabs for glory will only serve to mask the longer term underlying trend, the move to OpenCL across the industry.

All that together with the 28nm hardware from Red and Green Teams will make this year somewhat turbulent methinks :)

Regards
Zy

the situation is not so dark as you're depicting! AMD is surely pushing a lot on their compiler development, but also something else can be done by the projects, if they have the time and the ability (Matt seems a talented programmer!)
Surely, in the case of milkyway, where the code seems not so difficult, if the compiler is not giving best results, the kernel compilation can be done offline and then you can distribute a "hand-tuned" IL code. This was also something done by Gipsel back in the days, with the first version of the milkyway gpu app (Brook was not so good in those days, history is back ;) )

Anyway, a push back of 50 % with an app absolutely not tuned to AMD hardware seems not so bad. Even if without distributing IL tuned code, I think that some opencl compiler revisions and some tuning will cut this to no more than 20% without much effort.

We shall see what we shall see ..... hard to read the runes on this in absolute terms, as you point out talented programmers can do much to alleviate situations in many cases.

The underlying message should not be ignored though, whatever is done in the short term (my definition of that is circa 18-24 months), long term software trend is OpenCL in all its glory - a new set of goal posts we have to get used to like it or not.

Gonna be fun :)

Regards
Zy

Try Sapphire Trixx - works on reference AMD cards
https://www.sapphireselectclub.com/ssc/TriXX/TriXX.aspx
Regards Zy

Zy, thanks for the link. Question: just tried this (4.1.0) on a box with 2 5850 cards (different brands) and it mostly seems to work except that whenever I set the custom fan profile for one card the other resets to automatic. Tried everything I can think of, any ideas? Afterburner 2.10 does not exhibit this problem.

None - sorry - its a generalised utility that covers essentials, a throw away vender bribe rofl - so its kind of as is. Cant think of any workaround it might have for that.

Regards
Zy

We shall see what we shall see ..... hard to read the runes on this in absolute terms, as you point out talented programmers can do much to alleviate situations in many cases.

The underlying message should not be ignored though, whatever is done in the short term (my definition of that is circa 18-24 months), long term software trend is OpenCL in all its glory - a new set of goal posts we have to get used to like it or not.

Gonna be fun :)

Regards
Zy

in 18-24 months I think many more surprises can happen.
C++ AMP & similars can deliver much, if they keep up to promises!