@SekeRob - thanks for the reply. I have been playing with this since the last time I posted and an still scratching my head. I am having a little trouble navigating the WCG website infrastructure so any advice you have on where to find things would be appreciated.

Is there an "easy" way to search out other users who may be running one of these Bulldozer/FX chips? It sure would be nice to get some sort of positive affirmation that other users of this CPU family are able to run GFAM without issue.

Since I last posted I have replaced memory and under clocked the system - not that the FX needs any help in under performing.

I also jumped the kernel up to a 3.2 version which has some FX optimizations.

None of this was of any help. Each and every GFAM task I run on this system goes inconclusive and then into the invalid bin.

I am able to run rosetta, poem, and other WCG tasks without issue - at the original speed or at the under clocked speed.

I still suspect a cumulitive "rounding error" using the 256 bit floating point vs the more industry standard 128 bit unit. I know, I can be pig headed.

Where can I look to see evidence on what the validators believe was wrong with the output produced by GFAM tasks run on this system? (the WCG website is very complete, but the two things I have yet to be able to find is how do you isolate the RAC for a single system which is useful when optimizing systems, and where do you go to look at debug or diagnostic information)

I can read a stack trace and a panic dump. I am also proficient in setting breakpoints so if there is a way to get the same task to run on multiple systems, I can halt the task and compare results.

Once again, any information you can provide would be appreciated.

BTW - I've had GFAM "disabled" in the profile for this system so there are no current tasks to peer at. However, if you can provide me with a "crowbar" to open the lid and examine the internal workings I will happily reenable GFAM to get a few samples.

Thanks

Lawrence - thank you for the response and the information. You are correct that Bulldozer's 256 bit FPU can run as two discrete 128 bit units. And it appears that updated compiler support is required to use it in the 256 bit mode. Further, AMD's FX supports Fused Multiply / Accumulate (FMA) operations while running in the 128 bit mode.

An FMA op feeds the result of a floating point multiply directly into the accumulators without rounding which increases both the speed and the precision of the operation. The FMA op seems to be much like what you would see in a GPU.

I have yet to determine when the FMA operation is used instead of the "traditional" method of doing a "multiply" followed by an "add"

I have been focused on the FPU because that is one of the big differences between the FX and previous processors.

I guess the next step for me will be to toss together a few "test" routines I can run on both the FX, one of my Phenom's and my Intel based Mac Pro.

It sure is nice being retired and having the time to do this...

I recall that during the DSFL and GFAM early betas, WUs for both Phenoms and Athlons were going inconclusive to invalid. Then the techs adjusted something in the validation, and ever since the next round of betas, my Phenom II (also running Ubuntu 10.04.3) has had no issue with DSFL and GFAM.

I'll be very interested to see how this unfolds with the Bulldozer chips.

@davyboy: Thank you - I was starting to question both my sanity and technical savvy. Now only my sanity is in question.

For those not familiar with AMD's Opteron 6274, it is is a fairly new offering from AMD for the server world. It implements the same Bulldozer microarchitecture used in AMD's FX chips.

I guess it is time to turn this one over to the admins/devs - however, if there is anything I can do to help resolve this issue, just ask. I "opened up" my bulldozer based system to GFAM for a short while this morning and by this evening there should be several "fresh" examples of this issue in case anyone wants to examine them in the light of day.

Once again DavyBoy - thanks.