If the primary reason why GPU's are so efficient is having lots of SIMD multi threaded cores (which seems to be not particularly interesting as a discovery), why is it considered such a novel breakthrough?
I would think of it as a breakthrough particularly because of the sheer computational capacity that GPUs provided over a standard CPU. GPUs are specialized hardware that do really well for single-instruction multiple-data situations, and nothing else. They don't execute singular non-SIMD instructions, they don't do branch prediction or speculative prediction, but they do a remarkable job at parallel tasks like image processing, ML training/inference and real-time graphics acceleration.
Please log in to leave a comment.