We already have top level chasis, able to hold up to 8 NVDIA Tesla K series cards, the most powerful aimed at scientific calculus.
This node, designed to be racked, offers a great density because it only used 3 U been able to work either as an individual machine or a cluster node.
GPU based HPC systems are more popular because they allow the performance whole cluster in a machine.
In fact, this solution might give us 70 Tflops in simple precision and 23 Tflops in doble precision, thanks to the lastest Tesla K80. Not long ago, this performances were needed a full rack cluster and were 5 more times expensive, regardless communications, space, cooling costs, etc.
To be able to manage all the process with only one machine, also reduces management complexities and the latency between the nodos.
Aplications as Amber14, Gromacs, NAMD, LSMS, Chroma, River flow 2D, all of them in the Scientific computing field, improving these real performance between 5 and 15 times GPU vs. CPU.