On 05/02/2013 07:41 PM, Vincent Diepeveen wrote: > Also utter useless it is, as the GPU has what is it 200GB/s bandwidth, > going back to 6GB/s cache coherent gpu/cpu (just taking a number i've > seen more there get claimed > that the bandwidth was in OpenCL) you lose roughly factor 30+ so > using such features is total nuts.
The vector unit (formerly the GPU) would still have the local caches, no? Also, the APUs these would be replacing already have physically unified access to the DDR3 memory pool, so there's no change in effective bandwidth. Some APUs would come with a GDDR5 memory pool, so memory bandwidth would be upgraded for the CPU. The big change is that the pools are not separately allocated. _______________________________________________ Beowulf mailing list, [email protected] sponsored by Penguin Computing To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
