4 ms·
> is data-parallel computing on a GPU much worse somehow? OpenCL is a very awkward way to do vector processing - everything is hard-coded to an abstract model
by sedachv 12y ago
> is data-parallel computing on a GPU much worse somehow?
OpenCL is a very awkward way to do vector processing - everything is hard-coded to an abstract model of a typical consumer GPU memory hierarchy. CUDA is even worse with a ton of versions all having different limitations according to what the Nvidia chips can do.
It's awkward to do a lot of SIMD tasks on GPUs. The Connection Machine was a general-purpose SIMD originally designed for parallel graph algorithms.
OpenCL looks like what the Connection Machine C* language might get macroexpanded into prior to compilation: http://people.csail.mit.edu/bradley/cm5docs/CStarProgrammingGuide.pdf http://people.csail.mit.edu/bradley/cm5docs/CStarProgramming...
- abecedarius 12y agoThanks. I've read Hillis's book but not studied the modern stuff; I'd gotten the impression the hardware was capable of about as much, and faster now, even if organized differently -- shared memory instead of a network -- and with less-pleasant languages.
- sedachv 12y agoThat was pretty much where Thinking Machines was headed before bankruptcy - their third model (CM5) was a SPARC cluster primarily for running Connection Machine Fortran and C*.