Intel® FPGA AI Suite: IP Reference Manual

ID 768974
Date 9/06/2023

A newer version of this document is available.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

Document Table of Contents

2.4.1. Architecture Description File Format for Instance Parameterization

The Intel® FPGA AI Suite IP has a highly configurable architecture. Configuring the design allows for different trade-offs between inference performance (throughput and latency) and utilization of FPGA resources (area). Configurations are specified through Architecture Description Files. The IP instances corresponding to these configurations can be compiled as part of an FPGA design into an FPGA bitstream.

The architecture determines how much FPGA area is consumed by the Intel® FPGA AI Suite IP and strongly affects the achieved inference fps and ease of timing closure.

Achieving the best performance of a given graph for a given FPGA area (or the smallest FPGA area for a given performance target) requires optimizing the architecture. The architecture optimization function of the Intel® FPGA AI Suite compiler is designed to produce good architectures for a given graph or set of graphs. For more details about the architecture optimization function of the compiler, refer to the Intel® FPGA AI Suite Compiler Reference Manual .

The Intel® FPGA AI Suite Architecture Description Files use the protobuf format and have a .arch file extension. While these files are human readable and editable, manually optimizing an architecture requires a deep knowledge of the Intel® FPGA AI Suite IP design and is not recommended.

You adjust some of the architecture parameters by hand, because the Architecture Optimizer does not modify them. For example, the optimizer does not modify the numerical precision (for example, fp16 or fp11) in the architecture file. Similarly, the optimizer does not modify details related to the AXI interfaces on the IP. In some case, you can improve performance of the resulting optimized architecture by the choice of these values.

When possible, modifying the graph or batch size might also result in performance improvements. For example, a graph that requires FP16 precision might have sufficient accuracy at FP11 or FP12 if a few extra layers are added. Reducing the internal precision enables a large memory and area reduction. Very small and fast graphs might achieve a higher performance on hardware by using a batch size that is greater than one.

The example_architectures/ directory includes an example that shows how to enable the hardware-accelerated softmax function.

The comment character in the .arch format is #.