Ottawa-based fabless semiconductor developer Blumind has closed $20 million CAD in Series A funding to commercialize what it describes as an analog chip built from scratch for artificial intelligence (AI) to process physical information like sound and vision rather than digital signals.
The all-equity round was co-led by Cycle Capital and the Crown corporation BDC Capital, with participation from return investors including Fusion Fund, Two Small Fish Ventures, and Real Ventures. Cycle Capital managing partner Claude Vachet and BDC partner Remi Fournier are joining Bluvision’s board as a result of the round, which the company said it is keeping open as it “engages in active discussions with strategic investors.”
“If we didn’t have a computer today … how would we build a computer architecture for neural networks?”
Founded in 2020 by CEO Niraj Mathur and CTO John Gosson, Blumind claims its analog chip architecture was designed to reduce the power consumption and latency of always-on AI applications.
The startup has raised $34 million CAD to date, including this latest round.
The new capital will enable the startup to develop a high-volume production version of its product, lay the groundwork for its next-generation of chips, and hire approximately 10 employees to its engineering and design teams, Mathur told BetaKit in an interview.
Mathur said that traditional digital processors, which rely on thousands of transistors to relay the ones and zeros of binary code, were built to support human-written sequential code and, while they have their use-cases, the neural networks of AI applications are shoehorned in, though these two modes of computing have “inherently very different workloads.”
“When [Gosson] devised his architecture, it was all on the basis of, ‘If we didn’t have a computer today, if we didn’t have CPUs and GPUs today, how would we build a computer architecture for neural networks?’” Mathur said, referring to the computer processors and graphic processing units that modern AI applications rely on for compute. “That’s the premise of the company.”
Blumind isn’t the only semiconductor startup to seek a better way to handle the unique workloads of AI. Toronto-based Untether AI claims to have developed what it calls a “chip architecture for neural net inference,” which it says can reduce the distance data must travel to the absolute minimum, eliminating the cost of moving data to processors. However, Mathur said that Blumind is different.
“We’re not what’s referred to as an accelerator, we don’t merely assist another processor in running the neural network,” Mathur said. “We run the whole neural network ourselves.”
Mathur claimed that Blumind is using transistors in “a different way,”: He said rather than a one or a zero, its chips can store eight-bit values in a single transistor device. Blumind’s analog processor is meant for on-device, physical AI applications, where a digital processor could draw 100 to 1000 times more power than a Blumind chip does to remain “always on” to continuously detect and process physical information like audio, video, and sensors, Mathur claimed.
Mathur said he sees applications on the horizon in products like smart glasses, VR headsets, robotics, or vehicles with accelerometers and tire pressure sensors. Fusion Fund partner Shane Wall echoed the vision in a statement, saying that physical AI use-cases are emerging in segments where Blumind’s “low-power, low-latency and low-cost semiconductors will be a critical ingredient.”
Mathur said that his startup is still in its early stages, but Blumind plans to get its first product into production next year.
“We are already working with several large, tier one customers to do [proof-of-concepts], validate our technology’s performance, power consumption, and really tee up for the high volume production next year.”
Feature image courtesy Blumind.