Hi Jonathan,
The biggest impact will be in multi-GPU simulations. PCI-E 3.0 is used for inter-GPU communication and is the current bottleneck there. The multi-GPU pmemd.cuda scaling is currently limited by PCI-E bandwidth.
Actually, currently the most efficient (i.e. GB of trajectory data per day) way to run simulations on systems with multiple GPUs is to run one job per GPU. In our group, we use this approach and run multiple simulations of each system we're studying.
In those machines, we've mainly used PCI-E 2.0 with AMD CPUs and motherboards. See
http://ambermd.org/gpus/recommended_hardware.htm#hardware
With our current usage strategy, the main reason we've chosen AMD is that they're currently a lot cheaper than Intel right now and the CPU is not heavily used in GPU-accelerated pmemd.cuda. The DIY system with 2 GPUs has served us well and we get full performance out of each GPU.
Ben Madej
Walker Molecular Dynamics Lab
________________________________________
From: Jonathan Gough [jonathan.d.gough.gmail.com]
Sent: Thursday, February 28, 2013 6:53 PM
To: AMBER Mailing List
Subject: [AMBER] PCI-E 2.0 vs 3.0 (AMD vs INTEL?)
While doing some window shopping I noticed that the current AMD chipset
doesn't support PCI-E 3.0 x16 while Intel does. Muy understanding is this
would/does have an impact on the communication speed between the GPU and
the MOBO, but would that (or does anyone know if it does) have an effect on
the overall simulation speed (GPU-intercommunication)?
Apparently ASUS recently launched a MOBO that has an onboard chip to
compensate or cover the distance(?) such that 3.0 bandwidth is possible.
That being said I wonder if and how that might interfere with calculations.
any thoughts?
JOnathan
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Feb 28 2013 - 22:30:02 PST
