My concern on this topic is how to scale GPU computing on terabytes/petabytes of data. Admittedly, GPUs provide high performance computing. This allows us to scale down to smaller MR clusters. But as scaling down, we also lose the speed of moving data through the network and the speed of bringing data into memory for processing. Your Fatnode seems to tackle this problem, though, I am not so much familiar with such advanced hardware.

However, I still think MapReduce frameworks such as Hadoop should use this technique in their sorting/shuffling to cut down the time spending on these operations. They should also provide Hadoop GPU APIs and route these fuctions to GPU-CPU machines in the cluster. Basically, enabling GPU computing in MapReduce should be configurable and also should hide away GPU part from users.