Analog chips find a new lease of life in artificial intelligence

The need for speed is a hot topic among the participants of this week artificial intelligence summit – Larger AI language models, faster chips, and greater bandwidth for AI devices to make accurate predictions.

But some hardware startups are taking a regressive approach to AI computing to counteract the ‘more is better’ approach. Companies including Innatera, Rain Neuromorphics and others are working to create silicone brains with analog circuits to mimic brain function.

The brain is analog in nature, taking in raw sensory data, and these chip makers are trying to recreate the way neurons and synapses in the brain function in traditional analog circuits.

Analog chips can be very good low-power sensors, especially for some audio and vision applications, said Kevin Krewell, an analyst at Tirias Research.

“An analog is a much closer representation of how the brain works by using distributed memory cells to carry the weight of neurons or in some other way to carry an analog weight,” Crewell said.

Artificial intelligence and machine learning are mostly based on digital chips located at the edge or in data centers. But there is a place for analog chips at the edge, such as in smartphones or cars, which need instant intelligence without sending data to the cloud, which are used to deliver AI services.

“We are not aiming to replace the entire AI pipeline,” said Sumit Kumar, CEO of Innatera Nanosystems BV, based in Rijkswijk, the Netherlands.

Innatera’s third-generation AI chip contains 256 neurons and 65,000 synapses, which doesn’t sound like much compared to the human brain, which has 86 billion neurons, and operates at about 20 watts. But Kumar said it is possible to create a fully connected redundant network on top, and the chip could run on coin cell batteries.

The chip is used by customers to power radar and audio applications, with performance rivaling other chips in the same class. The goal of the slide is to incorporate low levels of machine learning and inference, which is considered a major challenge for artificial intelligence among presentation participants.

“What we’re trying to do is, what we realize is that when the data goes from a sensor to the cloud, it’s actually transformed in multiple stages by different types of AI. What we often see is customers processing low-level sensor data in the cloud,” Kumar said. , which is completely unnecessary.

The Inatera The chip takes the information from the sensor, which is converted into pins, and the input content is encoded exactly when these spikes occur.

“That’s exactly what’s going on in your mind. When you hear something, there’s… a little hair [cells] in your ear, which detects virtually every frequency band and what energy is within that band. And that hair [cells] It will vibrate, producing ripples, which then enter the rest of the auditory cortex. “Essentially, we’re following the exact same principle,” Kumar said.

According to this principle, inside the neurons of the brain there are calcium ions and low-sodium ions, and these concentrations change over time. The Intera chip replicates the same kind of behavior using currents.

“We measure how much current goes into and out of neurons. This is how we mimic the brain,” Kumar said.

The idea is not to disrupt the current flow of AI into the cloud, but to replace the current crop of AI chips on the edge that are unable to make on-device decisions. The chip also reduces the process of converting analog signals to digital.

“You can’t really translate an analog signal over a long distance because you’re actually getting degradation. We avoid that by converting that analog signal to surge,” Kumar said.

The basis of artificial intelligence today is based on simulating the work of neurons in the brain using digital chips and techniques, which has been very successful. Building on the advances in Moore’s Law, these digital circuits and networks are getting bigger and faster.

But the analog has its problems. For example, it is difficult to get consistency across analog chips with calibration issues such as drift,

“Analog circuits and memory cells don’t scale like digital circuits. Most of the time, analog eventually has to be converted to digital to interact with the rest of the system,” Crewell said.

The concept of neural chips is certainly not new. Companies like Intel and IBM are developing brain-inspired chips, and universities are developing their own versions using analog circuits. Intel and others have raised awareness of the difference between neural-shaped chips and traditional AI, but startups have felt the need to push their products forward as demands for AI computing and energy efficiency grow at an unsustainable pace.

Another AI chip company, rain nerves He said the brain simulation chip will be used in particle accelerators at Argonne National Laboratory.

In a presentation at the AI ​​Summit, the company didn’t give many details about how the chip would be used, but the company’s CEO, Gordon Wilson, said the chip would act like silicon brains that would help the research lab study and draw conclusions about particle collisions.

The silicon brain will provide on-device intelligence to guard against sensor drift, which can cause faulty data to be fed to AI systems. The concept of sensor drift is similar to typical drift in AI, where bad data fed into a learning model can cause an AI system to drift off course.

Wilson claimed that on-device chip capabilities are more energy efficient than AI in the cloud.

“You need the ability to learn on the go. You need the ability to train and tune against this sensor drift to keep this system performing,” Wilson said.

Wilson said that the first iteration of Rain’s chip “won’t look fundamentally different from … other analog or chip-mix”. But she will have the ability to learn, which will unlock more value.

Wilson referred to different types of memory, such as memristor circuits, that provide the ability to learn. Memristors have been in development since the 1960s, and HP (which later became HPE) followed suit for use in a massive computer called The Machine, but the technology is still new.

Memristor acts as a memory resistor. It is a resistor that can adjust its resistance. It’s used as a synthetic brace, Wilson said. In the brain, the synapses do not have to be perfect, and the requirements will be different for Rain’s diary.

Venture capitalist Sam Altman, known for his work in artificial intelligence as CEO of OpenAI, invested $25 million in Rain Neuromorphics earlier this year.

Leave a Comment