On Wed, 2011-10-05 at 22:42 +0200, Andrew Holway wrote: > Hi all, > > We are just in the process of buying a lot of the new AMD interlargo > boxes and I think my bosses are quite sold on the idea of providing > one to the Debian project for testing. We use Debian extensively and > are seeing that HPC performance, specifically floating point, is not > so great on the newer equipment.
In this chip, each pair of cores shares FPU resources. Whenever a core executes a 256-bit floating-point vector operation, it has to borrow resources from its neighbour. I'm no expert, but I would think they aren't the best choice for HPC. > Compared to the last step (magnycours) floating point performance is > roughly half on the new processors. > > Do you think access to a new box would improve this? While we could probably make good use of a new fast machine, I doubt that we could use them to improve floating point performance. That would normally be done by upstream developers working on compilers and numerical libraries. Ben. -- Ben Hutchings Once a job is fouled up, anything done to improve it makes it worse.
signature.asc
Description: This is a digitally signed message part
