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A newer version of this document is available.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

7.2. Top-Level Design Considerations

I/O constraints

In order to get the maximum performance from register retiming, wrap the top-level in a register ring and remove the following constraints from your .sdc file:

  • set_input_delay
  • set_output_delay

These constraints model external delay outside of the block. For the purposes of analyzing the effect of design optimizations, use all the available slack within the block. This technique helps maximize performance at the module level. Replace these constraints when moving to full chip timing closure.

Resets

If you remove reset generation from the design, provide a replacement signal by direct connection to an input pin of your design. This configuration may affect the retiming capabilities in Intel® Hyperflex™ architecture FPGAs. Add two pipeline stages to your reset signal. This technique allows the Compiler to optimize between the reset input and the first level of registers.

Special Blocks

Retiming does not automatically change some components. Some examples are DSP and M20K blocks. In order to achieve higher performance through retiming, manually recompile these blocks. Look for the following conditions:

  • DSPs: Watch the pipelining depth. More pipeline stages results in a faster design. If the logic levels in a DSP block limits retiming, add more pipeline stages.
  • M20Ks: Retiming relies heavily on the presence of registers to move logic. With M20K blocks, you can help the Compiler by registering the logic memory twice:
    • Once inside the M20K block directly
    • Once in the fabric, at the pins of the block

Register the Block

Register all inputs and all outputs of your block. This register ring mimics driving the block when embedded in the full design. The ring also avoids the retiming restriction with registers connected to inputs or outputs. The Compiler can now retime the first and last level of registers more realistically.