[Beowulf] What happened to AMD GPU?

[Prentice Bisbal]

M,

You're reading his original e-mail wrong. He didn't say AMD GPUs had a 
30-40% advantage over NVidia. He's asking IF people would switch to AMD 
if that were true.

Personally, I think it would have to be much more than 30-40%. The next 
NVidia card would probably provide that kind of improvement, so most 
NVidia owners would just wait for the next generation of NVidia product 
to come out instead of switching to another processor and port their 
code. How much better? I'd say at least 100%. You can't just have better 
performance, you need an overwhelming performance advantage to convince 
people who are already using CUDA to switch. 30-40% isn't going to do it.

If you can't do that, you have to go after new users who haven't started 
using CUDA yet, provide them with a value alternative to NVidia. Since a 
lot of public codes already support CUDA, you'd have to go after users 
writing their own code, preferably at the start of a project who haven't 
already invested in  CUDA.

Prentice

On 03/04/2015 02:33 PM, Massimiliano Fatica wrote:
> Looking at the list:
> FirePro s9150 has a score of 2.89/2.99 while the K40c has a score of 
> 2.98/3.15.
> Where is the 30-40% advantage?
>
> M
>
> On Wed, Mar 4, 2015 at 11:26 AM, Michael Di Domenico 
> <mdidomenico4 at gmail.com <mailto:mdidomenico4 at gmail.com>> wrote:
>
>     On Wed, Mar 4, 2015 at 2:07 PM, C Bergström
>     <cbergstrom at pathscale.com <mailto:cbergstrom at pathscale.com>> wrote:
>     > Frankly, it's visible we still aren't the best performance on some
>     > codes.. Those issues basically fall into two camps and we're working
>     > on it.
>     > -----------
>     > Having said this..
>     > Would anyone build an AMD GPU cluster if
>     > 1) The GPU was 30-40% faster than a K40
>     > 2) It cost less
>     > 3) OpenACC/OpenMP4/CUDA programming models were high performance
>     and supported
>     > 4) More oncard ram
>     > -----------
>
>     if the AMD's showed a 30-40% increase in performance over a k40, i'd
>     buy them.  but our lock in with CUDA would make it difficult...
>     _______________________________________________
>     Beowulf mailing list, Beowulf at beowulf.org
>     <mailto:Beowulf at beowulf.org> sponsored by Penguin Computing
>     To change your subscription (digest mode or unsubscribe) visit
>     http://www.beowulf.org/mailman/listinfo/beowulf
>
>
>
>
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.beowulf.org/pipermail/beowulf/attachments/20150304/0c1c49c5/attachment-0001.html>