Timely transmission of sensory data and control instructions is crucial between function modules of autonomous car systems. Socket-based message transmission mechanisms, e.g. LCM and IPC, are the de facto standard for this purpose because of their easy deployment and adaptability for distributed environment. However, they can no longer meet the increasing demands of high bandwidth for sharing large data packets, i.e. images and point clouds, among perception processing modules and the decision making module. The main reason is that socket-based mechanisms have to divide the large data packet into smaller data packets, which introduces extra costs for packing and unpacking, therefore results in high latency and low bandwidth. This paper proposes a new shared-memory-based inter-process message transmission mechanism. The shared memory segment is composed of a super block and several data blocks. The circular queues are deployed for fast scheduling data receiving and delivering. Meanwhile, the atomic operations are applied to improve the overall performance. Experimental results show that the average transmission delay of data packets of size 3MB is 2.5ms, which is drastically lower than LCM's 12ms and is also better than Sharedmem_transport's 3.9ms. The maximum throughput of this method is 1.1GB/s, which is much greater than LCM's 180MB/s and Sharedmem_transport's 600MB/s.