GPUs are already important compute engines in the HPC top 500, but GPU evolution has also made them a great fit for scale workloads like deep learning and inference. Recent increasing complexity of neural nets has created an exponential computation demand, requiring the use of GPUs for real time and high throughput inference / prediction, not just training. We’ll talk about GPU architectural evolution, changing neural net computation demands, and development flow from training to inference deployment as a utility in the datacenter. We’ll also cover data from large scale GPU deployments, and platform evolution for utility use, where predicting ExaFlop/s (that’s 10^18) capability is no longer a crazy number.