AI: IBM BlueGene/L top supercomputer, was: Re: [extropy-chat] Re: Life in Biosphere 1

[BillK]

On 6/23/05, Mike Lorrey wrote:
> What is interesting is that the end size of BGL is to be 131,000 or so
> processors. The current gigaflops per processor is ~2.1 at 62,000
> processors. When they are done, the gigaflops per processor will be
> 2.75. Is this from using ever faster processors as it progresses, or
> are they getting more work out of each processor the bigger the neural net?
> 

The processors are PowerPC 440 700 MHz (2.8 GFlops). 
It currently has 65536 processors. That's where the *theoretical*
maximum comes from.
2.8 x 131072 = 367 teraflops. 

The latest Linpack figures are:
Rpeak (GFlops):	183500
Rmax (GFlops):	136800

If you divide the Rpeak, 183500 / 65536 = 2.7999 GFlops.

Some of the experts are worrying about whether the mathematical
Linpack test is the proper test to use for supercomputers. It's not
that IBM have designed Blue Gene/L specifically to do well on the
Linpack test. It is just that their machine is designed for highly
threaded scientific work, which happens to be what Linpack is.
(Basically it measures floating-point processor performance).
Needless to say IBM and Intel like Linpack because they show up well on it.

Cray and AMD prefer the HPC Challenge Benchmark tests.
<http://news.taborcommunications.com/msgget.jsp?mid=400709>
<http://icl.cs.utk.edu/hpcc/index.html.php>

which also measure things like sustainable memory bandwidth,
communication bandwidth and random access integer memory updates.
 
Horses for courses really. 
Different applications, different benchmarks, different machines. 

BillK