Re: All CUDA devices are in prohibited mode, of compute capability 1.0, or otherwise unusable.

From: Kenno Vanommeslaeghe (kvanomme_at_rx.umaryland.edu)
Date: Thu Aug 15 2013 - 22:25:07 CDT

- In the normal consumer market, Nvidia does not sell complete graphics
cards; they only sell GPU chips. The third parties you listed build
complete graphics cards around these chips following Nvidia's
specifications. The different cards you buy may differ in memory size,
memory bus width, memory speed, memory quality and cooling. Memory size,
bus width and speed are part of the card's specifications so can easily
and objectively be compared (however, these properties are typically the
same for the same chip, especially for the higher-end cards). Conversely,
although some memory is known to make more errors than others, in the case
of graphics cards, it is almost impossible to get information on the
memory quality because when the card is used for graphics, a rare
transient error is usually imperceptible. So on that front, it's a
complete gamble. As for the cooling, you want a card that stays cool under
load so that the GPU doesn't make too many errors, and if you're gonna
build a cluster, you'll also want to see to it that that the cooling fans
won't all start failing in a few months, because replacing hundreds of
fans can be time-consuming. Luckily, the target audience of these cards is
quite focused on cooling, so you can get a lot of good information on that
front from reviewers on sites like newegg.com (I'm picking that particular
site because the quality of their reviews is a bit better than average).

- at the lower end of the range, Nvidia has been known to market different
chips with the same label. For consumers, they are roughly equivalent in
performance, but for GPGPU computing, significant differences may exist.
On this front, Wikipedia is one of the more informative resources:
http://en.wikipedia.org/wiki/Comparison_of_Nvidia_graphics_processing_units#GeForce_600_Series

- that said, even though the computing power of the higher-end cards is
almost exactly linear with the price, it is more relevant to look at the
price/performance of a compute node as a whole, and on that front, the
higher-end cards are clearly a better deal.

- *if* your workload requires substantial double-precision GPU performance
(which I *think* is the case for NAMD), the GeForce GTX Titan is supposed
to be miles ahead of any other Nvidia offering in the normal consumer
market, because it is the only consumer card on which double precision is
*not* disabled on a majority of the GPU cores.

On 08/15/2013 06:26 PM, Lucas wrote:
> 2013/8/15 Norman Geist <norman.geist_at_uni-greifswald.de
> <mailto:norman.geist_at_uni-greifswald.de>>
>
> The Ti has about 30% better performance, similar factor for the
> pricing. Same for the 700 series models. The pricing almost scales
> perfectly with the performance, except of the faster memory in the
> newer series. So at least it doesn’t really matter.
>
>
> Another thing that confuses me is that not only there are countless models
> to choose from, but actually different manufacturers for the same model
> too - the GTX 660 Ti model for example is available from EVGA, ASUS, MSI
> and Gigabyte, with prices ranging from $239 to $312. I've been reading the
> reviews comparing them, but most seem to be directed to the gaming
> community, who may have different needs when compared to us (it seems most
> of the difference is about "overclocking", which may be a good thing for a
> gamer but perhaps not for someone who actually needs consistency through
> much longer running processes). I know this is going off-topic, but I'd be
> thankful for more advice about it.
>
> Best regards,
> Lucas

This archive was generated by hypermail 2.1.6 : Tue Dec 31 2013 - 23:23:35 CST