On 21/03/2011 18:44, Lew wrote:
GPUs aren't just graphics processors. They're optimized for
caculations that are useful in graphics, like vector operations (four
32-bit doubles per 128-bit register!), multiply-add, bit-blt and
things like that that are useful for all kinds of things.
So a general-purpose CPU can ask a specialty one to calculate, say,
the product of a scalar and a 64-by-64 matrix and get substantial
performance boosts.
I think that when CUDA and friends came out PC power jumped an order or
two of magnitude (in theory). A home machine with 1TFLOPS became easy to
do. So I guess Moore's Law kind of spiked at that point.
Still we are factors of billions away from any realistic fundamental
computing limits. The Human brain does around 1PFLOPS/W and I doubt if
it's particularly well optimized.