Artificial intelligence is conquering the world and if you’re thinking we’re talking about software, you’d be correct. Relatively speaking, that’s the easy part. AI also requires a new generation of computer chips, and GrAI Matter Labs is designing those chips … brain-mimicking chips that give AI that near-human neural-networking ability to “think.”
This cutting-edge technology enables AI and ushers in new applications and possibilities.
“We're not doing the AI that sits in a server somewhere. We're doing the AI that fits in the tiniest devices,” says Menno Lindwer, vice president of Intellectual Property and Silicon at GrAI Matter Labs.
This is one of the High Tech Campus Eindhoven-based companies radically reshaping the future. Technically speaking, GrAI Matter Labs was founded and headquartered in Paris but now has a global presence with offices in Eindhoven and Silicon Valley. Lindwer and Atul Sinha, one of the company’s founders, located space on High Tech Campus Eindhoven. Both have spent most of their working lives on Campus and are quick to recognize HTCE as the place In Europe to find tech talent that can do what they do.
GrAI Matter Labs designs ultra-low power, fully programmable neuromorphic chips for machine learning. The team has built a completely new chip architecture, leading to great gains in efficiency and quality of output while decreasing latency and power consumption.
GrAI Matter Labs’ engineers and physicists have approached the problem in a novel way. Other companies optimize neural network accelerator chips to process just the machine learning part of AI. The problem is these can’t replace the original chip. Both chips are still needed. So, while the faster chip might accelerate AI by orders of magnitude, overall, the setup will only be marginally better in power or performance.
GrAI Matter Labs’ newest chip, GrAI VIP, is different.
“You don't need the original host processor to scale down the images and interpret them,” says Lindwer. “Our AI engine is self-contained and has everything to make it work in a standalone manner. It will run the sensor data capture, the pre- and post-processing and the whole neural network without ever having to talk to a host.”
There is no need for a separate CPU. The smart chips process the AI autonomously and this is how the team at GrAI Matter Labs makes chips they demonstrate to be orders of magnitude faster.
This processing autonomy also has benefits for privacy. Because the chip doesn’t have to process the data using another machine or server, all the data stays inside the device and never leaves.
The company has shown that by connecting sensors to its chip, it can autonomously process audio and video. GrAI Matter Lab’s tiny chip can, for example, be trained to take music from any source and split out any of the constituent stems, e.g., the vocals from the rest of the music without any degradation.
These are the initial steps the company is making towards a future where sensory inputs – audio, visual, tactile – can be manipulated with AI, rather than merely interpreted.
The team has proven its chip can be built into applications, such as smart robots that can track a person, speakers that process music and even medical devices that recognize heartbeats and other real-time audio inputs.
In fact, they are working on a project with Philips to test the use of AI in MRI machines. This project will also test the use of AI-enabled deep brain stimulation. In short, the AI will stimulate and tune electrodes on the outside or inside of the brain.
This may seem like the stuff of science fiction films, but it is backed by data from the company’s partners.
GrAI Matter Labs has achieved all this in seven years, founded in 2016 with a theory that bio-inspired processing could be applied to make calculations like the brain.
“DARPA funded us to test this idea. We proved it could be done, although not in an economically viable way,” says Lindwer. “But it provided a brain-inspired direction which eventually resulted in our NeuronFlow™ technology, to a large extent, the backbone of our current chip.”
The company received funding in 2017 to develop its first chip, GrAI One. In two years, the GrAI Matter Labs team designed the first version, built the software tools to program it and proved it could run neural networks. After another round of funding and another two years of development, in March 2022, the company launched its latest chip, GrAI VIP.
“GrAI VIP is a much-improved version and it’s only marginally larger. It still fits easily on your fingernail and consumes very little power,” Lindwer says.
While the HTCE office moves the technology forward, the Silicon Valley location is key in finding larger customers and funding. The U.S.-based team is on the lookout for large, forward-thinking companies with a willingness to work with small, yet ambitious, startups such as theirs.
“We want to build our chip for everyone, for all consumers,” explains Lindwer. “We come from backgrounds in Philips, Intel and Qualcomm. It’s in our DNA to build chips that sell in the millions. We know what it takes to bring chips to the consumer market.”
GrAI Matter Labs has shown it is not only a company looking toward the future but is also inventing the future.
“That’s why we don't have much trouble finding people to join us,” says Lindwer. “Our vision of the future is very much rooted in reality and we can show what we've done so far. People find it interesting and very challenging.”