Re: [AMBER] Multiple CPU and single GPU doubt

From: Ross Walker <>
Date: Sun, 3 Jul 2011 04:59:50 -0700

Hi Fabricio,

To add some to what Jason said. If you compare the performance gain in AMBER
from using the GPU to NAMD you'll see that AMBER's acceleration is much
higher, especially for a single GPU. This is because the entire calculation
is done on the GPU and communication back to the CPU is only done when you
need to do I/O. The cost of such CPU to GPU and back communication is very
expensive and so one wants to avoid it as much as possible. Hence trying to
use extra CPUs would actually just result in a net loss of performance due
to the need to keep exchanging forces over the PCI bus. NAMD sees
performance gain because they are focusing on 1 million atom+ systems where
the wall time time per step is long enough that you do not pay such a high
price since the computation to communication ratio is better. We are
focusing on the 20K to 800K atom range where communication is more costly
per step since you typically are running many more steps in a given time.
For this reason the extra CPUs do not help you with performance.

This may change with PCI-GEN3 although the real gain there will be PCI to
PCI device transfers without involving CPU memory so I suspect it still will
not be very beneficial to try to use the CPUs until the entire assortment
all share the same memory space.

However, you do not have to let the remaining cpus (cores) go to waste. You
can always use the unused cores to run a separate calculation and it should
not affect the GPU performance too much. I.e. if you have 8 cores and are
running 2 GPU runs (or a MPI job on 2 GPUs) you can use the remaining 6
cores to run a CPU (pmemd.MPI) job.

All the best

> -----Original Message-----
> From: Fabrício Bracht []
> Sent: Saturday, July 02, 2011 3:36 PM
> To: AMBER Mailing List
> Subject: [AMBER] Multiple CPU and single GPU doubt
> Having my compilation problems solved. I have one doubt left. Is there
> a way to run amber on multiple processors while still using
> accelerated GPU pmemd. I asking this because with namd, one can
> achieve greater performance while using multiple processors and the
> gpu for a CUDA-accelerated calculation.
> Thank you
> Fabrício
> _______________________________________________
> AMBER mailing list

AMBER mailing list
Received on Sun Jul 03 2011 - 05:30:04 PDT
Custom Search