Comments on: All NVIDIA, all the time: what’s up? http://insidehpc.com/2009/10/01/nvidia-news-surge-explanation/ HPC News Without the Noise for Supercomputing Professionals | insideHPC Sun, 09 Jun 2013 01:54:13 +0000 hourly 1 http://wordpress.org/?v=3.1.1 By: John West http://insidehpc.com/2009/10/01/nvidia-news-surge-explanation/#comment-185306 John West Thu, 01 Oct 2009 20:43:34 +0000 http://insidehpc.com/?p=7474#comment-185306 Brian - I hope to be able to dig into those details as well. I've sent emails...just waiting on a response. Hopefully I'll have helpful information to share. Brian – I hope to be able to dig into those details as well. I’ve sent emails…just waiting on a response. Hopefully I’ll have helpful information to share.

]]> By: Brian http://insidehpc.com/2009/10/01/nvidia-news-surge-explanation/#comment-185304 Brian Thu, 01 Oct 2009 20:17:04 +0000 http://insidehpc.com/?p=7474#comment-185304 I agree that this is interesting news, and I'm happy to see it here.. my only complaint (and, really, I shouldn't use that word) is that there's not more information on ORNL's planned 20 PF system. And I don't mean details on the system architecture, but rather on the procurement / research process... the Fermi chips, being equipped with essentially 256 double-precision cores, are obviously quite nice. But did ORNL get access to simulators? Early designs? Did they run any of their applications? Etc. See, the trouble with GPUs in the past has been the effort to use them - and unless you could get by with single-precision FP, that effort was seldom worth it since it often involved re-writing things (as opposed to, say, simply recompiling). Now the new chips offer a substantial bump to double-precision FP, so is ORNL planning on investing lots of effort in rewriting applications, or are they expecting the toolset to be substantially better, more or less allowing recompiles to take advantage of the hardware? Or are they targetting the single-precision crowd? I'd love to know more about that. What happened behind the scenes. I'm going to bug people at SC09, but I'd love to hear more if you can find out! (My place of work has yet to come around to the idea of investing even tiny amounts of their considerable funds into GPUs or other accelerators, but if I can say, "Hey, look - ORNL is doing this, that and whatever!", it helps make the case to shuffle some resources towards this. And maybe, just maybe, we can climb a little closer to the fore-front of HPC technologies.) I agree that this is interesting news, and I’m happy to see it here.. my only complaint (and, really, I shouldn’t use that word) is that there’s not more information on ORNL’s planned 20 PF system. And I don’t mean details on the system architecture, but rather on the procurement / research process… the Fermi chips, being equipped with essentially 256 double-precision cores, are obviously quite nice. But did ORNL get access to simulators? Early designs? Did they run any of their applications? Etc.

See, the trouble with GPUs in the past has been the effort to use them – and unless you could get by with single-precision FP, that effort was seldom worth it since it often involved re-writing things (as opposed to, say, simply recompiling). Now the new chips offer a substantial bump to double-precision FP, so is ORNL planning on investing lots of effort in rewriting applications, or are they expecting the toolset to be substantially better, more or less allowing recompiles to take advantage of the hardware? Or are they targetting the single-precision crowd? I’d love to know more about that. What happened behind the scenes. I’m going to bug people at SC09, but I’d love to hear more if you can find out!

(My place of work has yet to come around to the idea of investing even tiny amounts of their considerable funds into GPUs or other accelerators, but if I can say, “Hey, look – ORNL is doing this, that and whatever!”, it helps make the case to shuffle some resources towards this. And maybe, just maybe, we can climb a little closer to the fore-front of HPC technologies.)

]]>