>All the "use graphics cards as vector daughter boards" (see Nvidia's >latest announcements for the nv8800 as well) methods are very interesting >for HPC, and in that aspect highly-parallel vector units (such as Cell, >such as any multi-pipeline floating-point vector shader graphics card) >create a revival of the "floating point coprocessor". It's the HPC hype of >the day. I'll believe it once such a box enters the top500.
Xen on the PS3 would be nice; run Solaris and the PS3 OS at the same time :-) Joking aside, I am a bit warry of the "graphics cards as vector processors"; these are generally low precisions units and there was little incentive to make sure all the computations were absolutely correct; graphics doesn't really care *that* much. Certainly those errors would not attract attention nor would they propagate. The code running on those GPUs was always tightly controlled; so bugs in the hardware were readily worked around; will compilers which know how to avoid such errors be available for those situations? (This doesn't apply fully to Cell which has general purpose units rather than the embedded GPUs.) So it could all end in tears or Cell has it competition cut out from nVidia and AMD. But if the GPU moves to the CPU, will that not just mean we now do graphics in the CPU again because we now have the memory bandwidth we couldn't have before? Casper _______________________________________________ opensolaris-discuss mailing list opensolaris-discuss@opensolaris.org
