<br><div class="gmail_quote">On Mon, Feb 23, 2009 at 5:27 AM, Mark Hahn <span dir="ltr"><<a href="mailto:hahn@mcmaster.ca">hahn@mcmaster.ca</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
<div class="Ih2E3d"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
AMD Opteron "Shangai" 6MB L3 : 90% of peak<br>
Intel "Nehalem" Core i7 : 95.2% of peak<br>
Intel Itanium 2 : 95.6% of peak<br>
</blockquote>
<br></div>
interesting numbers, and Goto's efforts are always respected.<br>
it would be valuable to understand why, though. I usually think<br>
of ia64 as being basically designed specifically to make this kind of hero-optimization possible. but I wonder how well each of these chips does on low-to-medium quality user code and current compilers.</blockquote><div>
</div><div>As I see it, plenty of bandwidth and an excellent choice of cache sizes, latency and architecture. AMD had done a good job with Barcelona but didn't follow up with further improvements(other than cache size) with Shangai. Intel seems to have it the sweet spot.</div>
<div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;"><div class="Ih2E3d"><br>
<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
It's a Core i7 but should give you an idea.<br>
</blockquote>
<br></div>
I suspect that lots of people are holding their breaths (and POs)<br>
waiting for Intel to release 2 (and 4?) socket nehalems. it has to be a bit scary for AMD, since they're getting leapfrogged in the area<br>
of their traditional strength. I don't have a sense for whether Goto's<br>
kind of tuning (beyond the ability of compilers, and relatively<br>
cache-friendly relative to flops) is terribly relevant to the market.</blockquote><div></div><div>It's even worse for AMD in some other areas. Take a look at this: <a href="http://it.anandtech.com/weblog/showpost.aspx?i=554">http://it.anandtech.com/weblog/showpost.aspx?i=554</a></div>
<div></div><div>Best regards,</div><div> Tiago Marques</div><div></div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;"><br>
<br>
on the topic of memory bandwidth:<br>
<a href="http://techreport.com/articles.x/16448" target="_blank">http://techreport.com/articles.x/16448</a><br>
makes it sound as if AMD knows they have a scaling problem with cache<br>
coherency, and have a solution queued. does anyone know whether nehalem<br>
already has a probe filter? or if AMD has mentioned anything about widening to a 3-4-dimm-per-socket memory interface?<br>
<br>
regards, mark hahn.<br>
</blockquote></div><br>