All News | Boards | Chips | Devices | Software | Archive | About | Contact | Subscribe
Follow LinuxGizmos:
Twitter Facebook Pinterest RSS feed
*   get email updates   *

Snapdragon 835 neural processing SDK targets Android and Linux gizmos

Jul 26, 2017 — by Eric Brown 2,003 views

Qualcomm’s Snapdragon Neural Processing Engine SDK for the Snapdragon 835 supports Caffe, Caffe2, or TensorFlow AI frameworks on Linux or Android targets.

In May 2016, Qualcomm announced its first deep learning software development kit, called the Snapdragon Neural Processing Engine for the Snapdragon 820 system-on-chip. Now, it’s releasing a more advanced SDK for the Snapdragon 835 (APQ8098).

Simplified Snapdragon 835 block diagram
(click image to enlarge)

The Snapdragon Neural Processing Engine (NPE) SDK supports accelerated deep neural network workloads on Linux- or Android-based mobile and edge devices built on the Snapdragon 835. Development tools tailored for the Ubuntu 14.04 desktop support Java apps for Android or native formats written for Android or Linux. The platform enables developers to run a trained AI model on the devices without requiring a cloud connection.

The Snapdragon NPE SDK for the 835 enables developers to implement “deep learning user experiences like style transfers and filters (augmented reality), scene detection, facial recognition, natural language understanding, object tracking and avoidance, gesturing, and text recognition,” says Qualcomm. Applications include mobile, automotive, healthcare, security, imaging, AR gear, drones, robotics, and smart IoT edge devices.


Instead of adding “yet another library of network layers,” the NPE SDK supports Caffe and Caffe2 deep learning frameworks, as well as Google’s TensorFlow, says Qualcomm. The TensorFlow support was enabled by Qualcomm’s integration of TensorFlow machine learning service support in the Snapdragon 835’s Hexagon 682 DSP.

Snapdragon NPE development workflow
(click image to enlarge)

Features included in the Snapdragon NPE SDK for the Snapdragon 835 include:

  • Android and Linux runtimes for neural network model execution
  • Acceleration support for Qualcomm Hexagon DSPs, Adreno GPUs, and Kryo CPUs
  • Support for models in Caffe, Caffe2, and TensorFlow formats
  • APIs for controlling loading, execution, and scheduling on the runtimes
  • Desktop tools for model conversion
  • Performance benchmark for bottleneck identification
  • Sample code and tutorials
  • HTML documentation

After designing and training a Snapdragon NPE model, the file is converted into a “.dlc” (Deep Learning Container) file required by the runtime. The conversion tool outputs conversion statistics, including information about unsupported or non-accelerated layers. The developer can then adjust the design the initial model accordingly.

The octa-core, 10nm FinFET fabricated Snapdragon 835 comprises two quad-core banks of Cortex-A73-like Kryo CPU cores. The higher-end cores can operate at up to 2.45GHz when run alone, and can operate at up to 2.2GHz when working with the second bank of four 1.9GHz low-power Kryo cores. The SoC’s Adreno 540 GPU is claimed to deliver up to 25 percent faster graphics rendering than the Adreno 530.

Other features include an improved Hexagon 682 DSP, dual 14-bit Spectra 180 ISPs, and a Qualcomm Aqstic audio codec. The SoC offers built-in support for WiFi-ac and -ad, Bluetooth 5.0, NFC, and high-end, Cat 16 uplink enabled X16 LTE.

Further information

The Snapdragon Neural Processing Engine (NPE) SDK for the Snapdragon 835 is available for free download. More information may be found on Qualcomm’s Snapdragon NPE product page.

(advertise here)

Print Friendly, PDF & Email

Please comment here...