Neuromorphic computing draws inspiration from the brain and Steven Brightfield, chief marketing officer of a Sydney-based startup BrainChipsays this makes it perfect for use in battery-powered devices that perform AI processing.
“The reason for this is evolution,” says Brightfield. “Our brain had an energy budget.” Likewise, BrainChip’s addressable market is limited by power. “You have a battery, and only as much energy comes out of the battery as can power the AI you’re using.”
Today, BrainChip announced that its chip design, Akida Pico, is now available. Akida Pico, developed for use in power-constrained devices, is a scaled-down, miniaturized version of the BrainChip Ideology design, introduced last year. Akida Pico consumes 1 milliwatt of power, or less depending on the application. The chip design is aimed at the cutting edge, which includes small user devices such as mobile phones, wearables and smart home appliances that typically have severe limitations in terms of power and wireless communication capabilities. Akida Pico joins the market of similar neuromorphic devices designed for the edge, such as Innate‘S T1 chipannounced earlier this year, and The one from SynSense Xylo, announced in July 2023.
Neuronal spiking saves energy
Neuromorphic computing devices mimic the nature of the brain’s spikes. Instead of traditional logic gates, the computational units, called “neurons”, send electrical impulses, called spikes,to communicate with each other. If a spike reaches a certain threshold when it hits another neuron, that one fires as well. Different neurons can create spikes independent of a global clock, resulting in highly parallel operation.
A particular strength of this approach is that energy is only consumed when peaks occur. In a normal deep learning model, each artificial neuron simply performs an operation on its inputs: it has no internal state. In a spiking neural network architecture, in addition to processing inputs, a neuron has an internal state. This means that the output can depend not only on the current inputs, but also on the history of past inputs Mike Davisdirector of the neuromorphic computing laboratory at Intel. These neurons can choose not to fire anything if, for example, the input has not changed sufficiently compared to previous inputs, thus saving energy.
“Where Neuromorphic really excels is in processing signal streams, when you can’t afford to wait to collect the entire data stream and then process it in a delayed, batch manner. It is suitable for a streaming and real-time mode of operation,” says Davies. Davies’ team recently published a result showing theirs Long chipThe power consumption of was one thousandth of using a GPU for streaming use cases.
Akida Pico includes its own neural processing engine, along with SRAM units for event processing and model weight storage, direct memory units for peak conversion and configuration, and optional peripherals. Brightfield says that in some devices, such as simple detectors, the chip can be used as a standalone device, without a microcontroller or any other external processing. For other use cases that require additional processing on the device, it can be combined with a microcontroller, a CPU, or any other processing unit.
BrainChip’s Akida Pico design includes a miniaturized version of the neuromorphic processing engine, suitable for small battery-powered devices.BrainChip
BrainChip has also worked to develop AI model architectures optimized for minimal power consumption in their device. They demonstrated their techniques with an application that detects key words in speech. This is useful for voice assistance like Amazon’s Alexa, which waits for the keywords “Hello, Alexa” to activate.
The BrainChip team used theirs recently developed model architecture to reduce power consumption to one-fifth of the power consumed by traditional models running on a conventional microprocessor, as demonstrated in their simulator. “I think Amazon spends $200 million a year on cloud computing services to wake up Alexa,” Brightfield says. “They do this using a microcontroller and a neural processing unit (NPU), and they still consume hundreds of milliwatts of power.” If BrainChip’s solution actually delivered the claimed power savings for each device, the effect would be significant.
In a second demonstration, they used a similar machine learning model to demonstrate audio noise reduction, for use in hearing aids or noise-canceling headphones.
To date, neuromorphic computers have not found widespread commercial uses, and it remains to be seen whether these miniature edge devices will take off, in part due to the reduced capabilities of such low-power AI applications. “If you’re at the smallest level of the neural network, there’s only a limited amount of magic you can bring to a problem,” says Intel’s Davis.
BrainChip’s Brightfield, however, hopes that application space will be available. “It could be the awakening of the conversation. It could be as simple as noise reduction in earphones, AR glasses, or hearing aids. These are all types of use cases that we think are being targeted. We also think there are use cases that we don’t know if anyone will invent.”
Articles from your site
Related articles on the web
#Brainlike #computers #face #ultimate #limit