Perhaps sounds like a use case for a streaming parallel ADC board into GPU memory? Maybe an FPGA can run ADCs and buffer into a FIFO and transfer over PCI in large, efficient frames?