Crossed the 1 PFLOP barrier now
18 Nov 2013 13:49:40 UTC
Topic 197275
(moderation:
Hallo!
Today E@H passed the first time since the competition the 1 PFLOP barrier. At the end of the competition we reached 1.050 PFLOP at maximum, but it dropped relative rapidly to disapionting 850 TFLOP. After a long valey the crunching speed of the project increased slowly but almost constantly and we can hope for further permanently growth.
The next steps in technology will be the 14nm structurwidth coming to market next year and the 10nm some years later. 7 or 8 nm structurwidth are under research/development. I think we will see this to the end of this decade. So the technology will give the basis for further growth at limited electrical power consumption. Good for science!
Kind regards and happy crunching
Martin
Crossed the 1 PFLOP barrier now
)
I am happy to be a small part of this. Distributed computing is very cool.
Hm, BOINCstats lists Einstein
)
Hm, BOINCstats lists Einstein RAC at about 479.5 TFlops with yesterday clocking in at ~530 TFLops...
Note that Berkeley at some time changed the GFlops formula from RAC/100 (the one which I remember from the very beginning of BOINC) to RAC/200, don't know exactly when that happened or why.
http://boinc.berkeley.edu/wiki/Computation_credit
So we may have to wait some time for the magic 1 PFlops barrier...
(I only recently stumbled across the "new" formula as I thought I finally crossed my personal TFlops barrier - turned out I was not :p )
Hallo FalconFly! According to
)
Hallo FalconFly!
According to the history that change took place sometime between Sept. 2008 and May 20010, but there is no reason said for this change. I refere to the E@H server status page and this figure is used also in official papers. Interesting discrepancy, maybe Bernd will give a statement to this.
Kind regards and happy crunching
Martin
Yeah, in this case it would
)
Yeah, in this case it would be interesting to hear what the Einstein Team itself thinks of that... After all these figures are an important measure of a projects' success.
(I would assume such a change would be thoroughly discussed by the Berkeley devs with all project admins - turns out that maybe wasn't the case)
RE: Yeah, in this case it
)
I attended the BOINC Workshop in London, in August 2010, and I remember the change being announced to all attending projects (Bernd was there too) in David Anderson's "end of term report".
It was a technical correction of a programming error - benchmark speeds (they had been used as the basis for credit up until that point) had been summed instead of averaged, or vice-versa - that had been discovered while re-checking the maths for CreditNew.
It still strikes me as odd that Flops are derived from credit, rather than the other way round.
Hm, in order to reverse the
)
Hm, in order to reverse the calculation, BOINC would need a very comprehensive benchmark suite regularly optimized for real-world project performance. IMHO they always shunned those efforts.
I also think in the past many project admins made it clear that they sometimes had no way of predicting the GLFops count of their distributed workunits due to their client design (that was during the first discussions of implementation of fixed credit allocations, where those were known)
I was working BOINC improvement suggestions for the Devs myself many years ago and at least back in those days, they.... let's say they weren't too cooperative even towards readily presented solutions and even outright rejected to see some issues as actual issues. I guess they had other priorities.
Anyway, off-topic, but ever since several of the "easy credit" projects established themself (yielding horrendous credit compared to other projects), RAC IMHO lost quite some credibility anyway.
Well, there are people around
)
Well, there are people around who have tried replacing BOINC's benchmark with a more scientifically sound one, and - with a few other bugfixes - verified that the result gives more consistent and plausible results under CreditNew. Possibly only theoretically so far - I don't think they've reached the stage of even trying to ask a project administrator to alpha-test the results yet.
Oh well, at least they're
)
Oh well, at least they're working on it.
I remember at one time - simply due to total granted credit numbers growing so high for the 1st time - the BOINC devs actually discussed just regularly cutting all already granted credit totals by 100 or 1000 (with the argument that new users would have a better chance to improve their ranks).
At least those plans didn't realize back then, that was a terrible idea ;)