Don't forget fluid simulation and other dynamic particles. Real flow can utilize hundreds of CPUs with very low memory overhead. I've always wanted to test out a Maximus setup, now maybe I could afford a phi.
On Nov 15, 2012 10:34 PM, " mylenium@mylenium.de" < mylenium@mylenium.de> wrote:
Not sure what the fuss is about. Your normal Xeon processor has 8MB of L3 caches and you don't store image buffers in that, either. From what little reading I have done on the Phi card, internally all memory management and syncing is handled by the processors themselves via ring buses and crossbar switches in groups of 2, 4 or 8 cores, but that only really includes the caches and the internal processing queue. Also by the time you would process data on the card it would long be pre-segmented and unlike graphics cards, there would be no persistent buffers requiring extensive amounts of memory. The rest would be handled in the system memory and the beauty of it is that since the Phi card uses native x86/SSE etc. commands, there is no need for any extra data conversions or explicitly shuffling stuff around beyond what your system already does, anyway. Of course on some level it still has similar requirements like CUDA in terms of parallelizing stuff, but as Greg said, the biggest issue for the card at this point will be limited PCI transfer speeds. That's why primarily it will make inroads in science first most likely as well as 3D renderers that already to data segmentation/ bucketing/ tiling and by comparison the processing outweighs the data transfers. I wouldn't expect it to be relevant for anything else soon. Where intel will take this technology is another question considering how they have been mucking around for years, but generally I would not expect them to follow the same road that CUDA has.
Mylenium
[Pour Mylène, ange sur terre]
-----------------------------------------
www.mylenium.de
my math says each CPU would get 161mb of RAM.... despite what intel says, writing code to work well in this environment would not really be much different than CUDA. you certainly could not run an OS + App + App data on each core. thats fantasy. you could probably write a renderer to fit into that space and have some image buffers.... but when you watch AE eat a couple gigs for large deep color renders, their model pretty much doesn't work. At best, you instead would need to MP render each frame, tiling it up and feeding each CPU a chunk of it. very different than what intel is pitching. you'd need more like 64gb or 128gb to really make it work they way they say.
S
So, in theory would a multi-threaded renderer like Mental Ray be able to utilize all those coprocessor cores? If so, I imagine you would need a lot of RAM in the machine.
|
