Effects of Custom Deep Learning Processor Parameters on Performance and Resource Utilization

Analyze how deep learning processor parameters affect deep learning network performance and bitstream resource utilization. Identify parameters that help improve performance and reduce resource utilization.

This table lists the deep learning processor parameters and their effects on performance and resource utilization.

Deep Learning Processor Parameter	Deep Learning Processor Module	Parameter Action	Effect on Performance	Effect on Resource Utilization
TargetFrequency	Base module	Increase target frequency.	Improves performance.	Marginal increase in lookup table (LUT) utilization.
ConvThreadNumber	conv	Increase thread number.	Improves performance.	Increases resource utilization.
InputMemorySize	conv	Increase input memory size.	Improves performance.	To learn how to optimize your input memory size, see InputMemorySize and OutputMemorySize Optimization.
OutputMemorySize	conv	Increase output memory size.	Improves performance.	To learn how to optimize your output memory size, see InputMemorySize and OutputMemorySize Optimization.
FeatureSizeLimit	conv	Increase feature size limit.	None. This option allows the support for a fully connected (FC) layer with a larger feature number.	Marginally increases resource utilization.
FCThreadNumber	fc	Increase thread number.	Improves performance.	Increases resource utilization.
InputMemorySize	fc	Increase input memory size.	None. This option allows the support for a FC layer with a larger output activation.	To learn how to optimize your input memory size, see InputMemorySize and OutputMemorySize Optimization.
OutputMemorySize	fc	Increase output memory size.	None. This option allows the support for a FC layer with a larger output activation.	To learn how to optimize your output memory size, see InputMemorySize and OutputMemorySize Optimization.
InputMemorySize	custom	Increase input memory size	Marginally increases performance.	Increases Block RAM (BRAM) resource utilization.
OutputMemorySize	custom	Increase output memory size	Marginally increases performance.	Increases Block RAM (BRAM) resource utilization.
ProcessorDataType	Top Level	Change data type to int8.	Improves performance. There could be a drop in accuracy.	Reduces resource utilization.