<html style="direction: ltr;">
<head>
<meta content="text/html; charset=ISO-8859-1"
http-equiv="Content-Type">
<style type="text/css">body p { margin-bottom: 0cm; margin-top: 0pt; } </style>
</head>
<body style="direction: ltr;"
bidimailui-detected-decoding-type="latin-charset" bgcolor="#FFFFFF"
text="#000000">
<div class="moz-cite-prefix">If you want to start playing with CUDA
and don't care about performance, but just want to get a feel for
what different coding paradigms do, anything Fermi and up is a
good start. NVIDIA started adding L1 and L2 caches from Fermi GPUs
(400 series). Don't expect the low end cards to rival good CPU
coding (basic GPU coding will probably still be a little faster
than basic CPU coding though, i.e single thread, no sse).<br>
<br>
The 4xx and 5xx series are Fermi (previous generation), most 600
series are Kepler, although not all of them, some of the lower end
ones are just re-branded Fermies.<br>
<br>
Interestingly, Kepler behaves more like the lower end Fermi cards
than the higher end ones (more core groups than instruction
schedulers, and thus requires instruction level parallelism).<br>
<br>
If you just want to start playing with CUDA and learn the effects
of different optimizations, the 430, 520 or 620 will do the job.
If you want to get performance, although not peak, at a somewhat
saner price, the 660 or 670 will also do a nice job (the 660
according to zap is around 1400-1800 nis).<br>
<br>
On 09/10/2012 12:56, Oleg Goldshmidt wrote:<br>
</div>
<blockquote cite="mid:877gqzvb33.fsf@goldshmidt.org" type="cite">
<pre wrap="">Micha Feigin <a class="moz-txt-link-rfc2396E" href="mailto:michf@post.tau.ac.il"><michf@post.tau.ac.il></a> writes:
</pre>
<blockquote type="cite">
<pre wrap="">If the CUDA code is older, you may be better getting the older 5xx
series NVIDIAs, as NVIDIA did some hardware changes that makes writing
code for kepler a bit harder. They also increased compute power but
not memory speed, so not all codes benefit from the increase. If you
do want to run CUDA/OpenCL with a worthwhile speed, I would go for the
gtx 570 (Fermi) or the gtx 670 (Kepler). If you skip CUDA you can go
for middle end cards.
</pre>
</blockquote>
<pre wrap="">
Whle we are on the CUDA topic... Assume I want to play with CUDA (no
problem with proprietary drivers). I just got a quote for GTX680 for
ILS2820. If I don't need absolutely peak performance but I don't want
to be stuck with "older" ways to write code, could you recommend other
options that will be cheaper?
Thanks,
</pre>
</blockquote>
<br>
</body>
</html>