Optimised software for DSPs accelerates AI development
Optimised software for Cadence Tensilica HiFi digital signal processors (DSPs) efficiently executes TensorFlow Lite for Microcontrollers, part of the TensorFlow end-to-end open-source platform for machine learning (ML) from Google.
The combination of edge-based ML running on the ultra-low power cores supports the increasing demand for pervasive intelligence in advanced audio, voice and sensing applications
The HiFi DSPs are the first DSPs to support TensorFlow Lite for Microcontrollers.
With the addition of optimised software support for TensorFlow Lite operators on the HiFi DSP cores, developers can now take full advantage of the TensorFlow platform.
This promotes rapid development of edge applications that use artificial intelligence (AI) and ML, removing the need for hand-coding the neural networks and resulting in faster time to market and higher performance.
Implementing AI at the edge, including devices that use voice and audio as a user interface, requires running the inference model on the device. This has many benefits:
- Eliminates the latency associated with sending data to a cloud service and waiting for the response to be sent back to the device
- Reduces power consumption associated with sending/receiving large amounts of data across a network
- Maintains privacy and minimizes security issues since the data never leaves the device
- Without cloud dependency, the device can be disconnected from the network and still operate
“Voice and audio AI applications are now mainstream, as voice-based user interfaces become more popular with consumers,” said Ian Nappier, product manager at Google. “TensorFlow Lite’s microcontroller software combined with optimised operators for the HiFi DSP makes developing and deploying innovative neural nets on low-power, memory-constrained audio DSPs easier than ever.”
“Our powerful crossover microcontroller, the i.MX RT600, includes a 600MHz Tensilica HiFi 4 DSP that delivers 4.8GMACs of performance,” explained Joe Yu, VP of microcontrollers at NXP Semiconductors. “This new MCU provides the optimal balance in power and performance. The compute power makes the i.MX RT600 ideal for our customers deploying voice, audio, and other neural network-based applications at the edge. Supporting the popular, end-to-end toolchain, TensorFlow, as well as other inferencing technologies, on the HiFi DSP will enable ML developers to take advantage of the compelling combination of compute and memory on this chip.
“Enabling ML at the edge saves power, protects privacy and greatly reduces latency,” observed Yipeng Liu, director of audio/voice IP at Cadence. “Tensilica HiFi DSPs are the most widely licensed DSPs for audio, voice and AI speech. Support for TensorFlow Lite for Microcontrollers enables our licensees to innovate with ML applications like keyword detection, audio scene detection, noise reduction and voice recognition, with the assurance that they can run in an extremely low-power footprint.”
Tensilica HiFi DSPs are part of the broad Cadence IP portfolio and support Cadence’s Intelligent System Design strategy, enabling pervasive intelligence.