MENU

Neuromorphic computing SoC aims at ultra-low-power machine intelligence

Neuromorphic computing SoC aims at ultra-low-power machine intelligence

Technology News |
By Rich Pell



The 55nm node based platform includes a Cortex-M3 processor from ARM, a Coolflux DSP from NXP Semiconductors, 12bit successive approximation register (SAR) analog-to-digital converter, power management, and support for analog blocks such as power on reset, brown-out detector, oscillators, temp sensor, crystal oscillator, and RC oscillators.

Eta Compute reckons to have a couple of advantages for its microcontroller platform. One is the ability to operate at voltages down to around 0.2V the other is its spiking neural network software.

The company was founded in 2015 on a mission to get to deep sub-threshold voltages – and therefore extreme low power – and to do so it turned to asynchronous logic. Eta has applied its delay-insensitive asynchronous logic (DIAL) technology to the Cortex-M3 and the Coolflux DSP to provide an extremely low-power processor that can consume almost nothing when waiting for an event but can scale up to conventiona performance at about 100MHz clock frequency when required. The platform has been proven in silicon in TSMC’s 55nm ULP process, the company said.

On its own that should give Eta Compute a best in class conventional microcontroller. But the second advantage is the development of spiking neural network algorithms that run on a soft artificial intelligence engine hosted on the Cortex-M3. This engagement with artificial intelligence has come about largely with the recruitment of Nara Srinivasa as chief technology officer in 2017.

“Our patented event driven processor architecture (DIAL) is combined with our fully customizable neuromorphic algorithms,” said Srinivasa, in a statement. “These will be the foundation of a diverse and wide-ranging set of applications that deliver machine intelligence to the network edge.” Prior to joining Eta Compute Srinivasa worked at Intel Labs including research into spiking neural network modelling and performance.

Next: Applications


The sort of applications Eta Compute is aiming its platform at include applications for limited speech recognition; photoplethysmography (PPG) algorithms in wearables, motion detaction and sensor fusion, said Paul Washkewicz, vice president of marketing. “We’re doing a few applications in house,” he said that was partly as a proof of concept. This might include recognizing the spoken numbers one to ten so that an always-on unit could initiate a dial-out to a telephone number.

Doing more complex applications would be something Eta Compute would collaborate on with its IP licensees, said Washkewicz. “In the future we might be able develop a software development environment.

Using software to run a neural network on a Cortex-M3 core goes against the industry trend of moving towards hardware accelerators; multiply rich hardware that operates on relatively low resolution data types. But Washkewicz argued that while that approach may be necessary for convolutional neural networks it is a different case when using spiking neural networks.

“With spiking neural networks you get a lot more done with less multiplies. And from a power consumption point of view we are best in class.” He added that spiking neural networks make use of sparser connectivity than concurrent neural networks.

Washkewicz added that the DSP is a big part of the solution. Many applications, including audio and imaging, can benefit from upfront signal processing, for noise reduction and speech pre-processing.

Nonetheless current neuromorphic computing R&D tends to look towards memory devices that encompasses the characteristics of a biological neuron. In that way it is conceived that a memory array could be a hardware implementation and close analog of a biologically inspired neural network. Presumably it is felt that hard neuromorphics confer a power-performance-area advantage.

Washkewicz said that for now Eta Compute is pursuing programmable silicon with a software stack on top and the DSP for preprocessing.

Next: Down to 0.2V


The Cortex-M3 processor can operate down to as low as 0.2V and consume as little power as 1microwatt. However, the design uses level shifters to maintain voltages for SRAMs, Washkewicz said. He said that typically clock frequency is reduced down to less than 100KHz at 0.2V and runs at 100MHz at above 1V.

For now, the platform is only available for manufacture through foundry TSMC. “We’ve got 90nm with lower leakage and 55nm for a bit higher performance,” said Washkewicz adding that Eta Compute has some activity at lower nodes but not necessarily with ARM cores.

Eta Compute offers a variety of licensing models designed including standard, perpetual, and turnkey NRE contracts.

www.etacompute.com
https://www.frontiersin.org/articles/10.3389/fnins.2018.00126/full
https://ieeexplore.ieee.org/document/8259423/ 

Related articles:
Eta raises funds for spiking neural networks
Low-power startup recruits former Intel chief scientist as CTO
Minima, ARM apply ‘real-time’ voltage scaling to Cortex-M3
Self-timed logic is Eta Compute’s low-power secret

If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :    eeNews on Google News

Share:

Linked Articles
10s