IBM Claims Breakthrough Energy-Efficient Algorithm 231
jitendraharlalka sends news of a claimed algorithmic breakthrough by IBM, though from the scant technical detail provided it's hard to tell exactly how important the development might be. IBM apparently presented its results yesterday at the Society for Industrial and Applied Mathematics conference in Seattle. The breathless press release begins: "IBM Research today unveiled a breakthrough method based on a mathematical algorithm that reduces the computational complexity, costs, and energy usage for analyzing the quality of massive amounts of data by two orders of magnitude. This new method will greatly help enterprises extract and use the data more quickly and efficiently to develop more accurate and predictive models. In a record-breaking experiment, IBM researchers used the fourth most powerful supercomputer in the world... to validate nine terabytes of data... in less than 20 minutes, without compromising accuracy. Ordinarily, using the same system, this would take more than a day. Additionally, the process used just one percent of the energy that would typically be required."
A more informative link (Score:4, Informative)
Here's a link with actual content on what the algorithm does:
http://www.hpcwire.com/features/IBM-Invents-Short-Cut-to-Assessing-Data-Quality-85427987.html
Re:I'd expect this (Score:2, Informative)
AH you are not seeing the real potential of this.
It took them 20 mins to do. Meaning they can do 70 more customers in one day. Charge 10% more for the same job. Poof 70x1.1xcost gross profit. Oh and it cost them 99% less power wise. Margin just went up by 99%. Meaning they can also undercut competitors in the field. Or resize the computer so it still takes a day and still sell by the same price point. So it just costs them 99% less to do powerwise and customers pay the same amount.
Wont cost them a thing. In fact I would be willing to bet they make even more. Wait until the MBA's are done spinning it.
Re:just trying to be relevant (Score:4, Informative)
Did you actually read the article?
Well, that's hard to do since there was no reference. But the guy seems to be talking about "Massively Parallel Low Cost Uncertainty Quantification". This is probably the same work as this:
http://portal.acm.org/citation.cfm?id=1645413.1645421 [acm.org]
The work has nothing to do with energy savings, it's just about a fast, approximate algorithm for a fairly common operation.
Common sense tells you that spending 20 minutes to do something takes less energy than taking up to a day doing the same thing.
My point exactly. The whole press release is analogous to saying that you save a lot of energy by compiling with "-O0" instead of "-O4".
Re:Clarification? (Score:5, Informative)
The conference proceedings are not online yet. So I am not sure. I could not even find the title of the talk on the conference web page
I know people who are at SIAM PP and they are all : "why are they talking about PP on slashdot ?". There was no major anouncement. I'll check the proceedings again next week, but I believe there is no major improvement. IBM is probably just trying to get some more light.
We can find the following IBM talks in yesterday page :
http://meetings.siam.org/sess/dsp_programsess.cfm?SESSIONCODE=9507 [siam.org]
The paper have the same author and name than this paper published last year :
http://portal.acm.org/citation.cfm?id=1645413.1645421 [acm.org]
So they are probable publishing an improvement on their 2009 work.
Here's the actual paper. (Score:5, Informative)
"Low cost high performance uncertainty quantification", full text available in PDF.
http://portal.acm.org/citation.cfm?id=1645421&coll=GUIDE&dl=GUIDE&CFID=77531079&CFTOKEN=42017699&ret=1#Fulltext [acm.org]
And, here's the abstract:
Uncertainty quantification in risk analysis has become a key
application. In this context, computing the diagonal of in-
verse covariance matrices is of paramount importance. Stan-
dard techniques, that employ matrix factorizations, incur a
cubic cost which quickly becomes intractable with the cur-
rent explosion of data sizes. In this work we reduce this
complexity to quadratic with the synergy of two algorithms
that gracefully complement each other and lead to a radi-
cally different approach. First, we turned to stochastic esti-
mation of the diagonal. This allowed us to cast the problem
as a linear system with a relatively small number of multiple
right hand sides. Second, for this linear system we developed
a novel, mixed precision, iterative refinement scheme, which
uses iterative solvers instead of matrix factorizations. We
demonstrate that the new framework not only achieves the
much needed quadratic cost but in addition offers excellent
opportunities for scaling at massively parallel environments.
We based our implementation on BLAS 3 kernels that en-
sure very high processor performance. We achieved a peak
performance of 730 TFlops on 72 BG/P racks, with a sus-
tained performance 73% of theoretical peak. We stress that
the techniques presented in this work are quite general and
applicable to several other important applications.
Re:Cool, they've "discovered" PostgreSQL. (Score:2, Informative)
ok, mr. anonymous, I work with all those wares, and the differences aren't *that* big, some percentage points in certain situations with certain hardware and certain transactions. and the very fastest way to run databases doesn't involve open source software, tpc.org will tell you all about that. it happens Oracle or DB2 on a big HP/UX or AIX is going to whoop open source ass with usual business needs on mid and large systems, but at huge cost and with vendor lock-in and limitations to customization and integration with other systems.
Re:I'd expect this (Score:2, Informative)
What they've done (Score:2, Informative)
For anyone who's interested in what these guys have done- the WHPCF'09 paper by Bekas and Fedulova (and going back a bit further, their 2007 paper by Bekas et al.) give the details.
In many statistical problems we end up with the problem of finding the diagonal entries of the inverse of a known symmetric and positive definite matrix A. For example, in linear regression the variances for the fitted parameters are found on the diagonal of inv(X'*X). When this matrix A is very large, the computation can be very expensive, since it requires O(N^3) time by conventional methods (Compute the Cholesky factorization of A and then use the Cholesky factors to solve for N right hand sides.)
Bekas et al. have developed a Monte Carlo approach that can give good (e.g. 2-3 digits of accuracy) estimates of the diagonal entries in inv(A) by using an interative method to approximately solve systems of linear equations involving A. The approximate iterative solutions take roughly O(N^2) time, and there are s of these systems to solve, where sN. Thus the computational complexity is lowered from O(N^3) to roughly O(N^2). Furthermore, you can solve these s systems of equations in parallel. Going one step further, you can do a lot of the computation in single precision, so it can be done on GPGPU's and other machines that don't do double precision floating point efficiently.