Intel® Hyperflex™ Architecture High-Performance Design Handbook

ID 683353
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A newer version of this document is available.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

2.1.5. Avoid Broadcast Signals

Avoid using broadcast signals whenever possible. Broadcast signals are high fan-out control nets that can create large latency differences between paths. Path latency differences complicate the Compiler's ability to find a suitable location for registers, resulting in unbalanced delay paths. Use pipelining to address this issue and duplicate registers to drive broadcast signals.

Broadcast signals travel a large distance to reach individual registers. Because those fan-out registers may be spread out in the floorplan, use manual register duplication to improve placement. The correct placement of pipeline stages has a significant impact on performance.

In Sub-Optimal Pipelining of Broadcast Signals, the yellow box highlights registers inserted in a module to help with timing. The block broadcasts the output to several transceiver channels. These extra registers may not improve timing sufficiently because the final register stage fans out to destinations over a wide area of the device.

Figure 4. Sub-Optimal Pipelining of Broadcast Signals
Optimal Pipelining of Broadcast Signals shows a better approach with duplicating the last pipeline register, and then placing a copy of the register in the destination module (the transceiver channels in this example). This method results in better placement and timing. The improvement occurs because each channel's pipeline register placement helps cover the distance between the last register stage in the yellow module, and the registers in the transceivers, as needed.
Figure 5. Optimal Pipelining of Broadcast Signals

In addition to duplicating the last pipeline register, you can apply the dont_merge synthesis attribute to avoid merging of the duplicate registers during synthesis, which eliminates any benefit.

The Compiler automatically adds pipeline stages and moves registers into Hyper-Registers, whenever possible. You can also use manual pipelining to drive even better placement result.