Blog post: "Unified GPU/SMP benchmarking scheme ..."

[Jesse_V]

Well my two cents is that I will be switching from a AMD 5770 to a Nvidia GTX 660 Ti . I've grown tired of waiting for AMD to help out F@H. The race has been won as far as I'm concerned. It's all about the science.

You'll need to make sure that the 660 TI is on the whitelist and we're still waiting for the Kepler core to be finished off enough to move out of beta testing.

[7im]

Done and done.

[JCM3500]

Hi, FAH Folks:

I recently purchased a "new" PC based on an AMD A8 with HD 6500 and HD6550D GPUs for all of $158....It's an ugly light lavender colored box, but runs well. I am running this 24/7 since 22 October 2012 on FAH (two GPU tasks, one uniprocessor task and one systray task). How long will you continue support for the AMD GPUs? I get a sense that NVIDIA may be the favored GPUs???

Thanks,

John

[Joe_H]

There is work being done on both AMD and nVidia support, much of the rest is speculation.

[7im]

This forum and this project is hardware agnostic. Nothing is shown any favoritism.

[JCM3500]

There is work being done on both AMD and nVidia support, much of the rest is speculation.

Many thanks

[bruce]

Hi, FAH Folks:

I recently purchased a "new" PC based on an AMD A8 with HD 6500 and HD6550D GPUs for all of $158....It's an ugly light lavender colored box, but runs well. I am running this 24/7 since 22 October 2012 on FAH (two GPU tasks, one uniprocessor task and one systray task). How long will you continue support for the AMD GPUs? I get a sense that NVIDIA may be the favored GPUs???

Thanks,

John

For a long time, Folding on AMD hardware used and interface called Brook. AMD discontinued support for Brook and now supports OpenCL. FAH has supported both interfaces but discontinued support for Brook when AMD said they no longer supported it. This means that only AMD GPUs HD5000 and above are supported. With a HD6500, you should be fine for a long time.

[Off topic, so don't answer me] Are there any more of those $158 computers available?

[alancabler]

Hi, FAH Folks:

I recently purchased a "new" PC based on an AMD A8 with HD 6500 and HD6550D GPUs for all of $158....It's an ugly light lavender colored box, but runs well. I am running this 24/7 since 22 October 2012 on FAH (two GPU tasks, one uniprocessor task and one systray task). How long will you continue support for the AMD GPUs? I get a sense that NVIDIA may be the favored GPUs???

Thanks,

John

For a long time, Folding on AMD hardware used and interface called Brook. AMD discontinued support for Brook and now supports OpenCL. FAH has supported both interfaces but discontinued support for Brook when AMD said they no longer supported it. This means that only AMD GPUs HD5000 and above are supported. With a HD6500, you should be fine for a long time.

[Off topic, so don't answer me] Are there any more of those $158 computers available?

"Off topic" notwithstanding, enquiring minds want to know!

Way to go John (JCM3500!!! Krylon is your friend.

[alancabler]

@ KWSN_PToT

I came down hard on you earlier and I do apologize.

When you think about it, the donors most adversely affected by the increased points of the new unified plan will be those donors who are making a significant contribution to the project with their BigAdv boxes, while running away from the rest of us in the points race, but even they aren't really affected that much. Anyone with the wherewithal to build/maintain a 100K+ ppd machine can certainly also accomplish some GPU folding to keep the peon folders like me at bay. The same thing applies at any level of contribution...

ATI users would only be adversely affected if support for their cards dried up entirely, as a "rising tide lifts all boats", so to speak. Ending ATI folding could adversely affect the company (AMD), which is already (perpetually?) working under financial constraints. Perhaps they could free up some programmer's time to get their wonderful cards working for f@h at/near nVidia levels. Such a move might quickly pay for itself and boost AMD's bottom line. Many fine donors have foregone what might have been a superior- for- gaming ATI card purchase and went with a lesser nVidia card, simply because of folding performance. Speculation about ATI ppd might be off base, as their architecture just might not be up to CUDA performance for math calculations.
People are like that, too. The smartest woman I've ever known can't consistently add 2+2 and get 4, but has magnificent and extraordinary insight concerning just about any relevant influence on mankind.

You pays your money and you takes your chances.

[KWSN_PToT]

No, speed is not everything. Science is everything. Speed is only one component.

As you noted, the addition of the QRB was a suppliment to the existing points system at the time. In the past, science = points, and it was linear. But the quickness of the return also increased the scientific value, and that added value went mostly uncredited. Before the QRB, that time factor was not accounted for by the points system.

The addition of the QRB was an attempt to assign points to represent the added scientific value of faster returns of work units. And the QRB, overall, has been helpful to the project. But in hindsight, some sort of loyalty bonus should have been given to existing folders before the QRB was started to normalize the previous contributions with the exponential nature of future contributions. A sort of previous points "conversion" to the new points values. That way 10 years of loyal folding couldn't be outstripped in only 10 days on the new points system. But like they say, hindsight is 20/20, foresight is not.

In my opinion, all existing GPU points should be "QRB" adjusted (1 time bonus) so that past contributions more closely match future contributions, going back to where SMP QRB started. That's only fair, right?! Unfortunately, Pande Group doesn't keep enough WU data to make that happen any more. Oh well. Fold on.

Thanks 7im, for the support, insight, and clarity.

I hope I have not created a fire-storm around this issue, and I do understand both sides of the argument.

But here's the simple math (and why I think GPU QRB is not such a good idea):
Given the same WU...
Let's say that we have two GPUs: one that is older and capable of 4:30 TPF, and another newer model that is capable of 3:05 TPF.
The older GPU will complete the WU in about 7:30:00, but the newer GPU will complete that same WU in about 5:08:20.
Therein is the inherent QRB - the newer GPU will be able to complete more WUs over the same [longer] period of time than the older one.
More WUs = More points. Is my 'math' wrong???

And to 7im's observation that PG doesn't keep enough stats to implement a 'consistency' bonus (which is reasonable) - well, there is EOC stats (and others).
EDIT: I realize that the external sites do not have access to PRCG or stats related to specific hardware - F@H does not publish that info for download.
I am constantly on EOC stats, and Jason Rabel does a remarkable job of keeping the data "neat & tidy". I'm fairly certain he has detailed stats going back to 2004.
Maybe EOC can be approached to provide some statistical/advisory help in this area (if PG would agree to consider 'consistency-of-contribution' as a factor in awarding points).

To the point of "speed is everything", I have two remarks for that...

ONE:
I understand that some projects are seemingly more urgent than others.
PG - with a little extra work - could do a better job with their assignment servers.
It has been my observation that when the client-app is attempting to get a new WU, it does/can transmit the "hardware info".
If some project is deemed to be a high-priority project requiring a minimum hardware level, then the assignment server(s) should not give one of those WUs to significantly slower hardware.
Additionally, just because BIGADV, etc are set on the client side, shouldn't automatically mean that the assignment server MUST honor that switch.
The assignment server should consider it a request that can be declined when necessary.

TWO:
The assertion that some professors or grad-students need results/data quickly in order to complete certain papers/theses - and are somehow entitled to preferential treatment - is a little hard to swallow.
PG and its project designers need to remember the scope and magnitude of the collective computing power that is being donated by hundreds of teams and many thousands of people -for free.
By rigging the system to effectively penalize contributors with older hardware, they must realize that many dedicated folders are already frustrated and feeling that F@H is becoming an exclusive club for the techno-elite.
The long term effect can be chilling. You risk losing a significant portion of long term contributors if you are hedging your bets on bolstering a smaller-but-better class of contributors.

In a not-too-indirect way, we contributors are subsidizing Stanford & PG.
Many of us have invested thousands of dollars to build DC/HPC systems.
On top of that, there is the recurring expense of electricity (etc) to keep them running.
A little patience & gratitude from those whose careers are significantly dependent on our contributions is warranted.
If the majority of contributors get "fed-up" and stop contributing, how would Stanford/PG make-up that computing-power shortfall?
Would corporate entities step in to help? Would the Stanford Board of Trustees and/or Alumni Associations pony-up to buy equivalent computing resources?

The vast majority of F@H contributors have an abiding belief in the promises & potential cures/remedies that can be derived from the research.
Far too many of us have lost loved-ones to Cancer, Alzheimer's, Parkinson's, etc.
We all want to help put an end to these horrid diseases and alleviate human suffering.
Stanford & PG contribute the brains, and we [collectively] contribute the brawn.
But when some contributors are being told (in a not-so-thinly-veiled way) that "your contributions are insufficient", it sets a bad tone.
That bad-tone can find resonance with people like me - who do have the technical sufficiency, but appreciate everyone's contribution.

Please note that these are my own opinions, and do not necessarily reflect the prevailing attitude of my team.
I am not - in any way - acting in a spokesperson capacity for the KWSN.

[k1wi]

It is probably a better idea to continue the discussion of qrb in one of the many qrb threads, rather than this one (as your points cover more than just this project or this new unified system)

Many of the points have been raised prior to the roll out to gpu. As yet no one has provided a better method of rewarding the quick return of work units. After all, faster computers more expensive per flop and quicker depreciating than cheaper computers.

[ArVee]

Is ver. 7 required to get the new 8057's? I've not got one yet, I see reports that it's released and my 3 machines with two 465's and one 550ti haven't come up with one and I'm wondering if it's because they run 6.41?

[k1wi]

Don't ask about beta projects in a non-beta forum

[P5-133XL]

Is ver. 7 required to get the new 8057's? I've not got one yet, I see reports that it's released and my 3 machines with two 465's and one 550ti haven't come up with one and I'm wondering if it's because they run 6.41?

p8057's are currently closed beta only. All detailed communications regarding them will be in the closed beta forums only.

[JimF]

By rigging the system to effectively penalize contributors with older hardware, they must realize that many dedicated folders are already frustrated and feeling that F@H is becoming an exclusive club for the techno-elite.
The long term effect can be chilling. You risk losing a significant portion of long term contributors if you are hedging your bets on bolstering a smaller-but-better class of contributors.

On the assumption that there are only so many points to go around (and if you just issue more to everyone it deflates their value, just like money), then which group would you want to provide the incentive to? I don't think you need to be a PhD at Stanford to figure out the right answer. Maybe you are right that just the higher output of the newer cards is enough of an incentive by itself, but the scientific value of the contribution can only be decided by PG.

I hope they get the right balance too, but just because a group of folders wants to get more points does not indicate to me that they have done it wrong.