Forums - Clarification on Available Hardware

2 posts / 0 new
or Register
Last post
Clarification on Available Hardware
kunalpaode
Join Date: 16 Jul 23
Posts: 4
Posted: Thu, 2024-04-18 22:58
Top

In the context of SNPE and QNN, is there a difference between the HTP and the DSP when it comes to hexagon architectures that are v68+?

I ask this because for the UDO code prodvided by the SNPE SDK, we can see HTP header files grabbed in DSP_v68 implementation of the Conv2D:

#include "HTP/core/simple_reg.h"

Another Question: 

When using snpe-dlc-quantize , I have noticed that when using the --enable_htp flag, it quantizes the model to 16-bit rather than 8 bit. Is this supposed to happen?

  • Up1
  • Down0
Re: Clarification on Available Hardware
christy2951hernandez
Join Date: 16 May 24
Posts: 1
Posted: Thu, 2024-05-16 05:07
Top
Hello,

In the context of Qualcomm Neural Processing SDK (SNPE) and Qualcomm Neural Network (QNN), the HTP (Hexagon Tensor Processor) and DSP (Digital Signal Processor) are both hardware components used for neural network processing on Qualcomm Hexagon architectures. While they share some similarities, there are also differences between them.
 
The Hexagon Tensor Processor (HTP) is specifically designed for tensor processing, which is common in neural network computations. It provides specialized hardware acceleration for operations such as convolutions, matrix multiplications, and other tensor operations commonly found in deep learning models.
 
The Digital Signal Processor (DSP) is a more general-purpose processor that can also be used for neural network computations. While it may not have the same level of specialized hardware acceleration for neural networks as the HTP, it can still be utilized for running neural network models efficiently.
 
Regarding your question about the differences between the HTP and DSP in hexagon architectures v68+, it's essential to consult the official documentation or reach out to Qualcomm's technical support for precise details. They can provide insights into any architectural differences, performance characteristics, and recommended use cases for each hardware component.
 
As for your second question about the snpe-dlc-quantize tool quantizing the model to 16-bit instead of 8-bit when using the --enable_htp flag, this behavior may be by design or could potentially be a bug. Quantizing a model to 16-bit instead of 8-bit can impact memory usage and computational performance, so it's worth verifying whether this behavior aligns with your expectations and requirements.
 
If you're unsure about the behavior of the snpe-dlc-quantize tool or if you believe it may be a bug, you can consider reaching out to Qualcomm's support channels or posting on their forums for clarification. They should be able to provide more information about the tool's behavior and any potential workarounds or solutions. 

Best Regards,
NYStateofHealth
 
  • Up0
  • Down0
or Register

Opinions expressed in the content posted here are the personal opinions of the original authors, and do not necessarily reflect those of Qualcomm Incorporated or its subsidiaries (“Qualcomm”). The content is provided for informational purposes only and is not meant to be an endorsement or representation by Qualcomm or any other party. This site may also provide links or references to non-Qualcomm sites and resources. Qualcomm makes no representations, warranties, or other commitments whatsoever about any non-Qualcomm sites or third-party resources that may be referenced, accessible from, or linked to this site.