When experiments are running the sensors generate about 1PB of data per second. They have to do multiple (I think four?) layers of filtering, including hardware level to get to actual manageable numbers.
It depends on which experiment. We call it trigger system. And it varies according to each experiment requirements and physics of interest. For example LHCb is doing now full trigger system on a software side (No hardware FPGA triggering) and mainly utilizing GPUs for that. That would be hare to achieve with the harsher conditions and requirements of CMS and ATLAS.
But yes at LHCb we discard about 97% of the data generated during collisions.
