| Index | Recent Threads | Unanswered Threads | Who's Active | Guidelines | Search |
 |
World Community Grid Forums
Category: Support
Forum: GPU Support Forum
Thread: Double - Single precision Flops |
| No member browsing this thread |
|
Thread Status: Active Total posts in this thread: 10
|
|
| Author |
|
|
Hypernova
Master Cruncher Audaces Fortuna Juvat ! Vaud - Switzerland Joined: Dec 16, 2008 Post Count: 1908 Status: Offline Project Badges: 
|
I was wondering what type of math are used for the HCC GPU WU that we are crunching. Double precision (I would suppose) or single precision floating point.
----------------------------------------The compute power of the GPUs are very different for these two types.  |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
You tell me how floating point the integer computations are of HCC for it to matter if single or double precision ;O
|
||
|
|
nanoprobe
Master Cruncher Classified Joined: Aug 29, 2008 Post Count: 2998 Status: Offline Project Badges: 
|
I was wondering what type of math are used for the HCC GPU WU that we are crunching. Double precision (I would suppose) or single precision floating point. The compute power of the GPUs are very different for these two types. They are single precision. If they were DP a vast majority of GPUs that are now running the HCC GPU app would not be able to run it because they either don't support DP or it has been so crippled by the manufacturer as to be useless.
In 1969 I took an oath to defend and protect the U S Constitution against all enemies, both foreign and Domestic. There was no expiration date.
  |
||
|
|
Hypernova
Master Cruncher Audaces Fortuna Juvat ! Vaud - Switzerland Joined: Dec 16, 2008 Post Count: 1908 Status: Offline Project Badges:
|
I was wondering what type of math are used for the HCC GPU WU that we are crunching. Double precision (I would suppose) or single precision floating point. The compute power of the GPUs are very different for these two types. They are single precision. If they were DP a vast majority of GPUs that are now running the HCC GPU app would not be able to run it because they either don't support DP or it has been so crippled by the manufacturer as to be useless. The reason I was asking is that in SP a fast 7970 board is rated around 3.2 Teraflops but in DP around 1.1 Teraflops. If I look at the fastest HD 7970 boards I have, I pull around 1.4 TFlops in terms of WCG rating considering that 1TFlops is 700'000 points per day. When looking at the GPU load it appears at 98-99% or fully used. There is a gap somewhere between these compute power ratings. |
||
|
|
branjo
Master Cruncher Slovakia Joined: Jun 29, 2012 Post Count: 1892 Status: Offline Project Badges:
|
According to http://boinc.berkeley.edu/wiki/Computation_credit 1 TeraFLOPS is equal to 1,400,000 WCG points per day (1 x 200,000 x 7).
----------------------------------------Cheers   Crunching@Home since January 13 2000. Shrubbing@Home since January 5 2006  |
||
|
|
Hypernova
Master Cruncher Audaces Fortuna Juvat ! Vaud - Switzerland Joined: Dec 16, 2008 Post Count: 1908 Status: Offline Project Badges:
|
I know that definition branjo and is in conflict but the reference here is 700'000 points = 1 TFlops.
----------------------------------------For me also this is a mystery.  |
||
|
|
branjo
Master Cruncher Slovakia Joined: Jun 29, 2012 Post Count: 1892 Status: Offline Project Badges:
|
OK, thanks Hypernova
---------------------------------------- Crunching@Home since January 13 2000. Shrubbing@Home since January 5 2006 |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
[think we discussed this before], it's not a mystery to those who look at the history... the reference computer of 1000 MFlops used to be equal to 100 Cobblestones, 1 Cobblestone = 1 credit, and one decided based an analysis, no it's not 100 cobblestone, but 200 cobblestones [so the reference project could continue do what it did, give 200 for a TFL rather than 100]. Of course, that is not true, just an interpretation what one could make of it. Meantime, when WCG adopted that system with server 700, it never came close to that number the reference site was awarding], so we are in the neither fish nor meat territory. You're free to switch to WCG soya recipe, whenever that gets published ;>)
|
||
|
|
branjo
Master Cruncher Slovakia Joined: Jun 29, 2012 Post Count: 1892 Status: Offline Project Badges:
|
[ot]Will stay with meat regardless if it is fish or other meat. Definitely not with soya
---------------------------------------- [/ot]Crunching@Home since January 13 2000. Shrubbing@Home since January 5 2006 [Edit 1 times, last edit by branjo at Dec 5, 2012 8:17:24 PM] |
||
|
|
Hypernova
Master Cruncher Audaces Fortuna Juvat ! Vaud - Switzerland Joined: Dec 16, 2008 Post Count: 1908 Status: Offline Project Badges:
|
[think we discussed this before], it's not a mystery to those who look at the history... the reference computer of 1000 MFlops used to be equal to 100 Cobblestones, 1 Cobblestone = 1 credit, and one decided based an analysis, no it's not 100 cobblestone, but 200 cobblestones [so the reference project could continue do what it did, give 200 for a TFL rather than 100]. Of course, that is not true, just an interpretation what one could make of it. Meantime, when WCG adopted that system with server 700, it never came close to that number the reference site was awarding], so we are in the neither fish nor meat territory. You're free to switch to WCG soya recipe, whenever that gets published ;>) You are right was discussed probably a few years ago but my memory did not remember the outcome. It is amazing that we have no real direct way to make a definite good enough measurement of how many MFlops or GFlops or TFlops in average we are crunching for a given project. There must be differences with projects as the WU's are different. The only value we have at WCG is the benchmark value that varies between 3.2 and 4.2 GFlops with my CPUs For my GPUs I have also the statement in the logfile that goes for 1.6 TFlops peak to over 10 TFlops peak for a top HD7970 board. We have the ratings for the CPUs and GPUs capability that the manufacturers give but these are probably the maximum values, and these do not match by far what is displayed in the logfile. By knowing how much we pull compared to the maximum the scientists could design better WU's to use the spare compute power available. Just a 10% improvement would be on the global grid make 70 TFlops of available power difference. I know it is a never ending discussion .............. any new ideas ? |
||
|
|
|