Folding Forum

Rattledagger wrote:Thanks for the info, even it doesn't make the Ati-stats much easier to explain. There was a 33% drop from 01.02.2012 until v7 was released 22.03.2012, so if wu-supply didn't start to dry-out for old Ati-4xxx-cards before later cards had a working client the drop seems a little strange.

How could that be? The HD 5000+ hardware wasn't supported until folks migrated to V7.

codysluder wrote:How could that be? The HD 5000+ hardware wasn't supported until folks migrated to V7.

Well, my Ati-5850 did run the v6-client, not sure how things was with the later cards, appart for at the minimum needing to override the clients lack of GPU-detection.

Oh, the supply of WU's for the 4000 and earlier cards was definitely drying up well before they finally ran out in June. Multiple posts from various folders in the months before that reported waits of hours, sometimes days between turning in a WU and getting a new one to process. The newer 5000 and 6000 series cards using core 16 were getting some work on the public beta releases of the V7 client before the full release of 7.1.52 in March. But not enough to offset the loss of the older cards' contributions in the stats. Someone who kept older versions of the GPUs.txt file might be able to say when 7000 series cards started showing up in it as supported.

The irony is that this chart was started to help quantify these discussions.
Unfortunately, by changing the methods of determining TFLOPS, this data will not be able to correlate in the future.

Regarding AMD GPU.

Since the beginning of the year, both AMD and NVIDIA were in steady decline.

1. newer cards were not supported..keepler and AMD GCN.
As people bought new cards and replaced gaming cards in their home rigs...they stopped FOLDING due to a lack of support.

2. SMP and ADV BIG gave far more points for far less power...people on the forums here and elsewhere were told that GPU was inefficient and actually slowed the SMP clients.
People followed the advice and only ran SMP, for more points, less power, less noise, and less heat.

3. Heat wave/drought in midwest USA...more GPU folders shut down because of air conditioning cost and just high temps...even if they never worried about points and efficiency before.


The final blow ...and the second precipitous decline of 70% in 2 months for AMD GPU was indeed most likely caused by driver conflict.

1. AMD FOLDING client had not been updated in some years.
2. AMD Drivers updated monthly and increased steady in game compliance and speed. On newer cards, over a year or more, these drivers increased frame rates and smoothness by up to 50% over initial drivers.
Seeing huge improvements in games, users subscribe to automatic driver updates
3. last two months the beta and even standard AMD drivers "broke" the existing FOLDING CLIENT.

Users choose the efficiency and benefits they saw and were accustomed to with updated drivers over FOLDING(which had yielded poor points and inefficient usage of the cards architecture for years anyway)

I think the graph and timeline fit my explanation much better than the phase out of 4000 series cards (though it might have been a contributing factor)

But looking at the bright side...the remaining 3000 AMD Folders have doubled their TFLOPS overnight.

Joe_H wrote:Oh, the supply of WU's for the 4000 and earlier cards was definitely drying up well before they finally ran out in June.

Supply drying-out between v7-release and June is understandable, but I was trying to find-out if something happened in February & March. While there was a thread in the v6-forums for 24.02. talking about a server being down, it atleast looks like server was up and running again later-on. The thread about lack of 4xxx-wu started 19.04. so is in the expected decreasing-period roughly a month after v7-release.

As for v7, these forums gives a very "good" impression there's nothing before v7.2.9 and since these clients wasn't released before in October 2012 it should be unlikely to find anything about things happening in February or March here... But, for some strange reason, the info about previous v7-clients is still hidden-away here.

Starting with the oldest posts for v7.1.52, immediately found a post from 29.03.2012 asking if the new GPU-core for Amd-GPU's would be released soon or not, and a link to another post revealed it wasn't released yet.

Meaning, until end of March, the only "released" Ati-core was the old BROOKS-core and this worked on atleast Ati-2xxx - 5xxx-cards...

So, the drop in February/March was because...

other factor of note:

Awachs went down rapidly during this period.

http://folding.extremeoverclocking.com/ ... =&t=181223

My understanding is that his rendering/computational infrastructure included GPU. I do not know if they are AMD or Nvidia.
I think he was donating spare capacity and as his business picked up, that spare capacity diminished...or was just too expensive to continue donating.

Rattledagger wrote:Starting with the oldest posts for v7.1.52, immediately found a post from 29.03.2012 asking if the new GPU-core for Amd-GPU's would be released soon or not, and a link to another post revealed it wasn't released yet.

That might be a post about an updated version, core 16 and projects using it were in beta by early 2011. An announcement that one of the projects went to Advanced and for V7 only was posted the end of March 2011. I have not dug through and searched for posts showing when the core 16 projects moved from advanced to full folding.

So, both cores were available at the time before V7.1.52 went full release from being the last public beta before the beta of 7.2.9. If I recall correctly, the Brooks core could also be used on the 5000 series and possibly the 6000 series ATI cards, but the emulation broke at some driver upgrade. They would have been competing with the older cards for a diminishing pool of WU's.

Finally, take a look again at the data in the accumulated stats on the Google doc. There are a lot of gaps in 2011 into early 2012. So the graph is an estimate at best based on limited data points. There is also a large spike in November 2011 for ATI, that ended about the beginning of February. Whatever caused that jump appears to have shut down then.

Joe_H wrote:Oh, the supply of WU's for the 4000 and earlier cards was definitely drying up well before they finally ran out in June. Multiple posts from various folders in the months before that reported waits of hours, sometimes days between turning in a WU and getting a new one to process.

Here's the thread with some 64 posts on the death watch in April/May/June for the final depletion of available Core 11 WUs for the old 4000 and previous ATI cards. Radeon pre-4999 Unsupported. (Very few WUs left).

There was a lot of speculation as when the final WUs would be available - and when the PG would finally pull the plug. The chatter finally died out when there was no definitive announcement for the final exit, so each user made their own decision on when they unplugged their old cards, and either stopped GPU folding, or got newer hardware. There are several posts in this 5-page thread of what tack individuals took (including me in describing my decision to pull the AGP bus HD4670 card and add a GT430 96-CUDA Core Fermi card for the pci-non Express bus.)

There's also this post from the same thread, where I had analyzed the March-April period documenting the transition from essentially full availability of the WUs to increasingly longer dead times.

I was just looking at the Stanford Client Stats Page and see that the hot links on the page are to woefully outdated GPU FAQ pages for both Nvidia and [edit]ATI - GPU2 and GPU3. The link for the ATI FAQ (see quote below) has lists of cards that have long had folding discontinued.

These links to obsolete and erroneous info can't be helping the cause of encouraging users to fold with very productive GPUs.

How can these pages get better FAQs to link to?

Basic Requirements:
•2xxx/3xxx/4xxx/5xxx ATI Video Card, or newer
•ATI Driver v8.1+, v8.3 or newer recommended, up to v9.2 (v9.3 not supported yet) (do not use OEM drivers)
•5xxx - v9.10 driver or newer, MUST use -forcegpu ati_r700 switch (w/ v6.23)
•AGP GPU aperture size in the BIOS must be set to 128 MB or larger
•Microsoft .NET Framework 2.0, with updates recommended
•Windows operating system, XP or newer

GreyWhiskers wrote:These links to obsolete and erroneous info can't be helping the cause of encouraging users to fold with very productive GPUs.

Ati, FAH & "very productive GPUs"...

Uhm, was under the impression one of the main reasons for Ati-users dumping FAH is because Ati-GPU's is very un-productive under FAH...


Oh, and thanks to everyone posting more info. It gives atleast some plausible reasons for the Ati-decline this year, until middle of October.

mdk777 wrote:The final blow ...and the second precipitous decline of 70% in 2 months for AMD GPU was indeed most likely caused by driver conflict.
(snip)
Seeing huge improvements in games, users subscribe to automatic driver updates
3. last two months the beta and even standard AMD drivers "broke" the existing FOLDING CLIENT.

By the stats, including the 10-day stats-delay, the 76% drop happened over only 14 days.
I've only seen one thread about the lower performance with the latest Amd-drivers and no reports of FAH refusing to crunch, but can of course have overlooked some other threads.

So, maybe it's just me, but I still finds it strange 76% of folders runs with driver auto-updates enabled, all of these detects over a few days the lower performance, for so silently just dropping FAH.

well whatever the reason, they are down.
I suspect they will stay that way until the issues discussed are resolved.

Not really any need to parse it any finer.

But... we may be underestimating the number of institutional cards that were in use. At the higest, AMD was only 20,000.
You take down a supercomputer using AMD GPU and you loose 5000 , cards fast.

I just checked the stats, and discovered that we regained the title as being faster than all distributed computing projects under BOINC combined, once again.

Jesse_V wrote:I just checked the stats, and discovered that we regained the title as being faster than all distributed computing projects under BOINC combined, once again.

The big problem with this is, you're trying to compare "non-native FAH-FLOPS" against "native BoincStats-FLOPS"...

STATS doubled total FLOPS again today.

http://fah-web.stanford.edu/cgi-bin/mai ... e=osstats2

Have to love the power of creative accounting.

From 500 to 1200 TFLOPS instantly....amazing progress.

We were investigating a major under-reporting (OSX stats said that there were no OSX machines). A few months ago, Christian Schwantes (FAH Team member) rewrote our main stats scripts and he fixed a very old bug in the stats reporting. That bug had a workaround in the osstats2 page, so now the workaround was failing and that led to finally getting osstats2 working.

After checking the db with the raw files from the work servers (it's an important check that we can know what the right answer is from this more laborious manual process of going through the WS logs), this result is now correct, fixing the previous under-reporting.

I'm double checking this issue w/Christian and if his audit of the new changes passes, I'll write a blog post about this.