Blumind startup pursues analog AI at the edge

Blumind startup pursues analog AI at the edge

Technology News |
By Peter Clarke

Blumind Inc. (Ottawa, Ontario, Canada) is a startup company that has developed an analog architecture for neural network inferencing at the edge.

The use of analog computation and storage for neural network inferencing has theoretical advantages in terms of power consumption but it also throws up challenges in terms of stability of stored weights and data against thermal variation, other variations and interferences.

Blumind was founded in 2020 by John Gosson (CTO) and Niraj Mathur (COO). The founding team was joined by Roger Levinson as CEO. Gosson has developed a patented analog semiconductor architecture to process neural networks which is the basis of the company’s chips. The company claims that the analog architecture consumes 100 to 1,000 lower energy than digital approaches.

The company taped out its first test chip in 2022. It is preparing to launch production chips in 2024. According to the company’s website: “AMPL is the first all-analog AI on advanced standard CMOS architected to fundamentally mitigate process, voltage, temperature and drift variations.”

Blumind has said that it intends to address ‘edge’ AI with battery-powered sensors and devices that are energy constrained. This will include the “always-on” type sensing that is used to minimize energy consumption in embedded systems. Examples include keyword detection, visual wake image and time series data analysis.

Two SoCs

The company has plans for two initial “calling card” chips: the BM110 SoC which is optimized for always-on audio and time-series data applications; and the BM201 SoC for always on video and image classification applications.

These are domains being pursued by another startup, Syntiant Inc. (see TinyML platform provider ports to Syntiant processor). It is notable that Syntiant started out with the intention of using analog-in-memory computation in flash memory, but reverted to a digital approach with which it has achieved success.

Reportedly Blumind is using analog storage in single transistors in conventional advanced CMOS manufacturing processes. The Blumind IP can be integrated into SoC and microcontroller products or made available as a chiplet known-good die for system-in-package integration.

The company also provides essential software to port applications to its hardware. Blumind makes use of conventional neural network development platforms such as PyTorch, TensorFlow, Caffe. After an application has been trained in the cloud using these platforms, Blumind translator software is used to map coefficients after quantization and compression into a weights file for the Blumind device. Weights to personalize the Blumind inference products are loaded at system power-up. The network can be retrained and updated with new weights as desired.

Blumind closed a second funding round in January 2023 for an undisclosed sum of money.

Related links and articles:

News articles:

TinyML platform provider ports to Syntiant processor

Energous, Syntiant team up for energy harvesting edge-AI

Renesas teams with Syntiant on voice-controlled visual AI

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


Linked Articles