| Lysis |
| (User) |
| Mon Aug 20 2007 05:22 PM |
|
GFDL super computer?
|
So I imagine the folks at the lab are gunning some serious Itanium-based superscaler monstrosity.
I am writing a paper on GPUs and parallel processing in general, and I was wondering if there is any practical application for Nvidia's massively parallel Tesla platform, or GPGPUs like it for long range forecasting?
A rackmount with only four units is capable of outputting about 2 teraflops of numbers.
I have been unable to find any meaty info on the very machines that power the forecasting tools that we talk about every day, and this is a field in which I am inordinately interested. :?:
| Lysis |
| (User) |
| Tue Aug 21 2007 06:03 PM |
|
|
|
Re: GFDL super computer?
|
As a follow up for anyone who is interested, I found some information in a wiki article of all places:
"The GFDL operates various-sized SGI Altix 3700 and 4700 supercomputers with a total processor count of over 5,000. Hardware updates occur on average, every 18 months.
Some impressive information from the Altix website, concerning the 4000 series:
" It supports up to 512 processors under one instance of Linux and as much as 128TB of globally shared memory."
"The GFDL has been utilizing High Performance Computing Systems for 50 years. A brief portion of that history can be found on the top 500 supercomputing list".
Linkage
Looking at the list, you can see that the highest ranking the gfdl ever achieved was 75th in November of 1996.
Currently it is ranked 112th.
I wonder how much of the model's effectiveness is capped by lack of funding and/or computing resources?
| Clark |
| (Meteorologist) |
| Tue Aug 21 2007 11:43 PM |
|
Re: GFDL super computer?
|
Re: limited performance -- to some degree yes, but to some degree no. We still have a lot of knowledge-based and data-based improvements to get at before significant improvements are going to come from improved computer power alone, IMO. Data ingestion, however, is one area where more computer power would help, mostly to allow for time-varying data to be ingested into the models.