From: J. Andrew Rogers (andrew@ceruleansystems.com)
Date: Tue Nov 17 2009 - 10:43:55 MST
On Nov 17, 2009, at 3:10 AM, Alejandro Dubrovsky wrote:
> On Tue, 2009-11-17 at 00:43 -0800, J. Andrew Rogers wrote:
>
>> Top 500 has zero -- repeat *zero* -- relevance to AGI. None. Zip.
>> Nada. It benchmarks a code that is completely orthogonal to
>> essentially all AGI workloads, and selects for systems that typical
>> AGI workloads would scale horribly on. Top 500 is scaling because it
>> increasingly excludes almost all useful workloads outside of an
>> extremely narrow application space.
>>
> That seems like way too strong a statement to me. Firstly, top 500
> gives you something like an estimate upper bound on the performance of
> almost any code you might write. It tells you how much you are giving
> up by not making your code more linpack-like.
Most codes can never be LINPACK-like, and hence why LINPACK is not considered a useful system metric for most HPC purposes. It is intrinsic to the algorithms being run. Almost the entire class of nominal AGI models do not describe a LINPACK-like code. The single exception might be whole-brain modeling and similar, since that obviously could be implemented as a molecular simulation. LINPACK is for cache-friendly numerical codes, which most codes are not.
AGI, regardless of model, is generally a graph-like code. These are the polar opposite of LINPACK-like codes and are bound by architectural aspects that would be irrelevant for LINPACK. Supercomputers optimized for graph-like codes look nothing like the XT5, and generally employ sub-gigahertz processors. The engineering reason sub-gigahertz processors are optimal for large-scale graph-friendly architectures is left as an exercise to the reader.
A graph-like code running on Jaguar will very likely be *slower* than a graph-like code running on a large desktop that uses a graph-friendly architecture despite the vast differences in nominal computing power. At the same time, graph-friendly systems are going to show very mediocre performance with LINPACK-like codes because they are optimized for the intrinsic operation rate of graph-like codes.
For Top 500 to be relevant to AGI, you would have to demonstrate that AGI does not employ any graph-like algorithms. I cannot think of an AGI model proposed on any AGI-related mailing list that has not been intrinsically graph-like.
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:05 MDT