Article ID: 000089133 Content Type: Product Information & Documentation Last Reviewed: 05/20/2022

Why Are Precision Values Different When Inferencing in GNA_SW_FP32 and GNA_HW Execution Modes?

BUILT IN - ARTICLE INTRO SECOND COMPONENT
Summary

Default precision values for GNA_SW_FP32 and GNA_HW execution modes

Description
  • Converted LSTM model into Intermediate Representation (IR).
  • Obtained two different precision values when inferencing in GNA_SW_FP32 and GNA_HW execution modes.
Resolution

When inferencing in GNA_SW_FP32 execution mode, parameters and calculations are substituted from low precision to floating point (FP32).

When inferencing in GNA_HW execution mode, the default precision value is I16.

Related Products

This article applies to 2 products