There are technologies which seem to pose a great promise, yet they always lack that final "push" that moves them from being a niche activity (sometimes of academic interest only) to a needed, mainstream, technology.I believe that DNN (Deep Neural Networks) on GPUs snowball has started going downhill, and it brings new life to GPGPU as part of it.

A year ago, during GTC 2014 keynote, NVIDIA CEO Jen-Hsung Huang dedicated a large part of his keynote to talk on Neural Network and Machine learning (link to video ishere)

During that conference, Dr. Ren Wu from Baidu showed his work on using DNN for image classification (article). In his session, starting from slide 24, he shows the most exciting aspect of DNN, which is why I believe it will bring new life to GPGPU.

The exciting thing about DNNs and GPGPU is that the problem (and technology) lends itself to two distinct markets where GPUs presence is strong - HPC and Mobile. The asymmetric nature of the DNN process denotes that the long, heavy, compute-intensive, training stage will run on a large cluster of powerful GPUs - while the trained network can easily run on mobile-class GPU. And here lies the key for claiming that this will become a killer application - as it combines the need for large GPU farms running DNNs learning phases for continuous tuning of the neural network with the use of GPGPU on any common smartphone (and in the future, wearable)

A great demonstration of that principle is the new Drive PX technology, presented by Mike Houston at CES 2015:https://www.youtube.com/watch?v=zsVsUvx8ieo . It showed how DNN serves as a great solution for ADAS (Advanced Driver Assistance System) and even paves the way for applying it as part of autonomous driving system. The followingslide from the PX deck shows the asymmetric aspect is applied here as well :

These usages add up - Image Classification/Recognition, ADAS, Deep Speech coming from Baidu (http://usa.baidu.com/tag/deep-speech/ ).Can't wait to see what will come at GTC 2015 ...

Yet, one thing is missing here - and forgive me NVIDIA, as you really pave the way... We need broad adoption to make it a success, and broad adoption has to come from using a cross-platform standard. In order to fulfill the promise and become the killer application, a DNN primitive library needs to be written in a standard such as OpenCL (or the new "kid" - Vulkan), and be optimized for various targets.