Blumind utilizes analog semiconductor architecture to facilitate AI inferencing

Peng Chen, DIGITIMES Asia, Taipei 0

Niraj Mathur is the co-founder and chief operating officer of Blumind. Credit: Blumind

Artificial intelligence development and adoption have encountered serious energy efficiency hurdles with current computer architectures. To solve this challenge, Canadian startup Blumind has invented an analog architecture for semiconductors that will process neural networks efficiently with extremely low power consumption.

Niraj Mathur, co-founder and chief operating officer of Blumind, said the company taped out its first chip last year and has received great results. It is building more test chips, gearing up for the production that aims to start in 2024.

A semiconductor architecture that has low power consumption

Founded in 2020, Blumind has a mission of proliferating AI inferencing with its semiconductor technology.

Mathur explained that AI training uses vast data sets to train the neural network. AI inferencing, which happens after training, is the function the neural network will perform for a user.

He took cameras using a trained neural network to detect a person's presence as an example. He said those cameras will perform the inferencing function – using the neural network to decide whether there is a person in the image it collects – over and over. The process will account for most of the AI workload as it needs to be always-on for the best user experience.

The industry has faced the obstacle of delivering sustainable and highly performant AI for the future, according to Mathur. To begin with, it will be challenging for the von Neumann computer architecture, which has been in use for decades, to fulfill the needs of machine learning and neural network processing in the future.

Mathur said the situation is compounded by the fact that Moore's law is plateauing and making it much more expensive to receive performance gains.

To help the industry deliver enhanced AI technology, John Gosson, Blumind's co-founder and chief technology officer, developed a patented analog semiconductor architecture to process neural networks. The founding team also includes Roger Levinson, chief executive officer. The three leaders possess over 70 years of semiconductor industry experience combined.

Mathur said the Blumind architecture operates like the human brain, which is trained over time and can use information gathered from biological sensors like the eyes, ears, and nose to make decisions.

He added that the analog architecture consumes 100 to 1,000 times lower energy, which is essential to the environment. Some industry estimates state that AI compute will use 90% of all the power generated by humanity today within 30 years if AI efficiencies stay the same as today.

"Clearly, that is not sustainable…we have to get more efficient with how we do machine learning," Mathur said.

He also said using analog compute to reduce power consumption is a great theory. But making the architecture manufacturable in high quantity volumes brings another set of challenges. The solution also needs to work reliably under environments like extreme temperature conditions. Blumind has addressed all these significant challenges in its technology.

Targeting edge AI space

Mathur said while Blumind's technology can apply to many fields, it is initially targeting the edge AI space, namely the battery-powered sensors and devices that are energy or thermally-constrained.

He said Blumind's technology can help cameras on a car or smart mobility device look for objects that the vehicle or device wants to avoid hitting. It could also enable an electrocardiogram (ECG) monitor or a glucose monitor to use a neural network to detect and process data the sensor collects in a patient's body with meager power, delivering many years of battery life or supporting energy harvesting approaches.

Augmented reality (AR) headsets can be another application. Mathur said AR is very compute-extensive and has tight power, latency and thermal constraints due to its small form factor and quick response times. Blumind's chips will facilitate the device in processing finger pose information efficiently without carrying a big battery.

Taiwan as a launchpad for Asian market

Mathur said Blumind closed the second funding round in January 2023 and will march toward a Series A round after making progress this year. He also said the company is engaged with a few high-profile, blue-chip customers to define its product. It hopes to have these industry leaders as lead customers when entering production.

Blumind is looking to connect with the semiconductor industry in Taiwan, which hosts leading companies like TSMC and UMC. Mathur said the company also sees the use case of its technology, including smart sensors and smart cities, being adopted more rapidly in Asia and Europe.

"We're also looking at Taiwan as a launchpad for the broader Asian market for our products in the future," he added.

Having been in touch with the Taiwanese semiconductor ecosystem, Mathur said he has been very impressed with the feedback he received. He will visit Taiwan to attend COMPUTEX 2023.

"I'm very, very excited to be there finally and meet people face to face," Mathur said.