The effect of adding a scalar D-cache to the Cray-4 vector processor

In the past, vector supercomputers achieved high performance with long arithmetic pipelines coupled with fast scalar processors. Processor speed has increased at a rate greater than memory speed. Indeed, current vector processors have cycle times far faster than the memories they are connected to. When compilers can predict memory access patterns, they vectorize computations and thereby hide the processor/memory disparity. When memory access patterns are not known until run-time, caches can pay large dividends. This paper studies the effects of adding a scalar data cache to a modern vector processor and shows some encouraging results.<>