Google's Cat-Recognizing Neural Network on GPUs

By Norman Chan

An artificial brain at 1/50th the cost.

Remember Google X's 16,000 CPU neural network from last year that was able to discover and recognize cats from YouTube videos? That computerized brain--composed of a billion-connection neural network--ran on a cluster of 1,000 computers that cost roughly $1 million. Well the team behind that project has just published a paper demonstrating that the same amount of computing power (meeting the requisite 1-billion connections) could be duplicated with $20,000 worth of over-the-counter consumer hardware, namely three servers each with two quad-core CPUs and four Nvidia GeForce GTX 680 GPUs. And 16 of these servers scale that neural network processing power to over 10 times the number of connections as the original CPU-based system.