[Developers] opencl newfmin example
ian.taylor at noaa.gov
Fri May 18 11:10:32 PDT 2012
I'm also assuming the different number of function calls with and without
the GPU is due to precision of the calculations. With the convergence
criterion set to 1.e-10, the values in the par file for bigmin are all
equal at 1.00000, you'd need to record the minimization path to understand
the differences better.
Amazon's EC2 "Quadruple Extra Large GPU Instances", are among their most
expensive offerings at US$2.10 per hour. They also sell the GPU
for $1299.00, so maybe the rental fee is not so bad. After you stop the
instance, you pay something like 10 cents per gigabyte per month in storage
if you want to keep it around. Here's the full price list:
For me the hardest part of using Amazon's cloud computing was getting
connected, which required two sets of public and private keys (one for
initial connection, one for the cluster itself). But for folks with
experience in those matters it should be easy. For something like stock
assessment, where you might have very high computing demands for a short
period of time, I think that Amazon's service could be really useful. But
it would require agencies to think about paying for computing resources in
a new way.
On Thu, May 17, 2012 at 7:48 PM, John Sibert <sibert at hawaii.edu> wrote:
> On 05/17/2012 01:46 PM, Ian Taylor wrote:
>> Here are run times for the bigmin model:
>> new newfmin with GPU: 0 minutes 22 seconds for 442 function evaluations.
>> new newfmin w/o GPU: 3 minutes, 37 seconds for 682 function evaluations.
>> old newfmin time (no GPU): 5 minutes, 44 seconds for 2119 function
> It obviously followed different paths to minima with and without the GPU.
> Were they the same minima? Is this caused by the precision issue Dave
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Developers