Article: Parallelism at HotPar 2010
By: AM (myname4rwt.delete@this.jee-male.com), August 5, 2010 3:57 am
Room: Moderated Discussions
Dean Kent (dkent@realworldtech.com) on 8/4/10 wrote:
---------------------------
>AM (myname4rwt@jee-male.com) on 8/4/10 wrote:
>---------------------------
>>...unless you can *really* back
>>your 100x-claims-are-crap statements by showing that every published work that reported
>>such speedups is not worth considering for being misleading, wrong etc.
>>
>
>First, I'll admit to not following this discussion closely, so I may have missed
>something, but from reading the back and forth of the last day or two, this is what it looks like from 20,000 feet...
>
>David proposes a theory (or hypothesis) that 100x claims are crap. He supports
>this theory by using examples of the claim (or claims) he is aware of.
No examples were shown. Even for a few specific papers.
>Someone
>disputes his claim... but not by giving an example, but by stating that David's
>theory is crap unless he can show that virtually every published work is incorrect.
>
>So, it seems to me you have your reasoning completely backwards. The way such
>things work, from my understanding, is that the theory is considered to be possible
>until someone provides a specific example of where it is wrong. That's all it takes
>- one example.
One example can be sufficient proof for e.g. "not all 100x speedup reports are correct", which is not the claim made.
At that point, there is no question or discussion about whether
>the theory is right. On the other hand, your approach can *never* prove 'truth',
>because it is always possible that there is some obscure published work, or an as-yet-unpublished
>work that David has not analyzed so that you can claim David could be 'wrong', and
>by extension, that this makes you 'right'.
>
>It is obvious that this discussion is taking on all the trappings of a pissing
>contest. It only requires that someone show the *one* example of a published work
>that disproves his theory, yet you try to put the onus on him for absolutely, indisputably,
>and in all other ways completely prove the truth of his statement by finding every
>possible published work. If that were the way the scientific method worked, nothing would *ever* get proven.
Actually a general claim needs a general proof, that's exactly what scientific approach would suggest. :)
>So, it appears this discussion isn't about getting to the truth, it is about 'being
>right' - or at least trying to cast someone else as 'being wrong' without actually doing the work.
We haven't seen any work from David that justifies his claims, but there's something I can suggest. In the link below there's a list of only 10 papers, all reporting 100x and greater speedups, and he is highly welcome to study them all and show what is misleading, bogus, wrong or otherwise fishy about the works and results reported therein.
And Mark Roulo is highly welcome to show that the *real* speedup is only 2.5x-5x as stated by him.
Fair enough I hope?
And, by the way, 10 papers is just a small selection, so strictly speaking finding some fishy stuff about all 10 unfortunately is not enough of a proof.
http://blogs.nvidia.com/ntersect/2010/06/gpus-are-only-up-to-14-times-faster-than-cpus-says-intel.html
>Just my observation, though as I said, I might be missing the actual point.
>
>Regards,
>Dean
---------------------------
>AM (myname4rwt@jee-male.com) on 8/4/10 wrote:
>---------------------------
>>...unless you can *really* back
>>your 100x-claims-are-crap statements by showing that every published work that reported
>>such speedups is not worth considering for being misleading, wrong etc.
>>
>
>First, I'll admit to not following this discussion closely, so I may have missed
>something, but from reading the back and forth of the last day or two, this is what it looks like from 20,000 feet...
>
>David proposes a theory (or hypothesis) that 100x claims are crap. He supports
>this theory by using examples of the claim (or claims) he is aware of.
No examples were shown. Even for a few specific papers.
>Someone
>disputes his claim... but not by giving an example, but by stating that David's
>theory is crap unless he can show that virtually every published work is incorrect.
>
>So, it seems to me you have your reasoning completely backwards. The way such
>things work, from my understanding, is that the theory is considered to be possible
>until someone provides a specific example of where it is wrong. That's all it takes
>- one example.
One example can be sufficient proof for e.g. "not all 100x speedup reports are correct", which is not the claim made.
At that point, there is no question or discussion about whether
>the theory is right. On the other hand, your approach can *never* prove 'truth',
>because it is always possible that there is some obscure published work, or an as-yet-unpublished
>work that David has not analyzed so that you can claim David could be 'wrong', and
>by extension, that this makes you 'right'.
>
>It is obvious that this discussion is taking on all the trappings of a pissing
>contest. It only requires that someone show the *one* example of a published work
>that disproves his theory, yet you try to put the onus on him for absolutely, indisputably,
>and in all other ways completely prove the truth of his statement by finding every
>possible published work. If that were the way the scientific method worked, nothing would *ever* get proven.
Actually a general claim needs a general proof, that's exactly what scientific approach would suggest. :)
>So, it appears this discussion isn't about getting to the truth, it is about 'being
>right' - or at least trying to cast someone else as 'being wrong' without actually doing the work.
We haven't seen any work from David that justifies his claims, but there's something I can suggest. In the link below there's a list of only 10 papers, all reporting 100x and greater speedups, and he is highly welcome to study them all and show what is misleading, bogus, wrong or otherwise fishy about the works and results reported therein.
And Mark Roulo is highly welcome to show that the *real* speedup is only 2.5x-5x as stated by him.
Fair enough I hope?
And, by the way, 10 papers is just a small selection, so strictly speaking finding some fishy stuff about all 10 unfortunately is not enough of a proof.
http://blogs.nvidia.com/ntersect/2010/06/gpus-are-only-up-to-14-times-faster-than-cpus-says-intel.html
>Just my observation, though as I said, I might be missing the actual point.
>
>Regards,
>Dean