This design does not do well at CPU parallelism for many workloads. Consider texture sampling in rasterization, to name just one example, in which shared memory is extremely important to get good performance.