> This work leaves much room for improvement, however. For example, the throughput and latency can be significantly improved by using GPUs and FPGAs to accelerate the computation.

https://www.microsoft.com/en-us/research/wp-content/uploads/...

> We demonstrate CryptoNets on the MNIST optical character recognition tasks. CryptoNets achieve 99% accuracy and can make more than 51000 predictions per hour on a single PC. Therefore, they allow high throughput, accurate, and private predictions.