Intel® Quartus® Prime Pro Edition User Guide: Platform Designer

ID 683609
Date 10/04/2021
Public

A newer version of this document is available. Customers should click here to go to the newest version.

Document Table of Contents

4.5.4.1.1. Acceptable Latency Increase

For a pipeline bridge, Platform Designer adds a cycle of latency for each pipeline option that is enabled. The buffering in the clock crossing bridge also adds latency. If you use a pipelined or burst host that posts many read transfers, the increase in latency does not impact performance significantly because the latency increase is very small compared to the length of the data transfer.

For example, if you use a pipelined read host such as a DMA controller to read data from a component with a fixed read latency of four clock cycles, but only perform a single word transfer, the overhead is three clock cycles out of the total of four. This is true when there is no additional pipeline latency in the interconnect. The read throughput is only 25%.

Figure 123. Low-Efficiency Read Transfer


However, if 100 words of data are transferred without interruptions, the overhead is three cycles out of the total of 103 clock cycles. This corresponds to a read efficiency of approximately 97% when there is no additional pipeline latency in the interconnect. Adding a pipeline bridge to this read path adds two extra clock cycles of latency. The transfer requires 105 cycles to complete, corresponding to an efficiency of approximately 94%. Although the efficiency decreased by 3%, adding the bridge may increase the fMAX by 5%. For example, if the clock frequency can be increased, the overall throughput would improve. As the number of words transferred increases, the efficiency increases to nearly 100%, whether or not a pipeline bridge is present.

Figure 124. High Efficiency Read Transfer