Re: GTX-660 Ti benchmark

From: Thomas Albers (talbers_at_binghamton.edu)
Date: Sun Sep 16 2012 - 14:47:35 CDT

Hello!

> It looks like at the moment the performance is roughly on par with a GTX
> 570 or M2070 card.
>
> I would suspect that with some optimization specific for these cards
> future releases might be much faster? Or is your sense that this is
> where they will be staying?

A couple more numbers - this is the F1ATPase benchmark with varying
numbers of cores:

 n(core) s/step
  1 0.248
  2 0.163
  3 0.146
  4 0.136

What we are seeing is that either the GPU is oversubscribed already with
two or three cores, or that there is a bottleneck somewhere.

> I guess one thing to note is that based on a rapid google-search, the
> H77 chipset would be a PCI 2.0, so the CUDA performance may be suffering
> from not being able to use the full card on account of bandwidth
> limitations, which could be alleviated with a PCI 3.0.

The southbridge itself provides several PCI-E lanes at 2.0 data rate,
however, the CPU (Ivy Bridge) provides 16 PCI-E 3.0 lanes, and the
graphics card was plugged in the 3.0 slot. Absent chipset issues, I
don't think that there is a bandwidth problem. (Does anyone have
suggestions how to check for that?)

Thomas

This archive was generated by hypermail 2.1.6 : Mon Dec 31 2012 - 23:22:04 CST