Frangipani
Regular
https://www.embedded.com/brainchips-akida-npu-redefining-ai-processing-with-event-based-architecture
01. Okt. 2024 / 15:55
BrainChip’s Akida NPU: Redefining AI Processing with Event-Based Architecture
Maurizio Di Paolo Emilio
6 min read
0
BrainChip has launched the Akida Pico, enabling the development of compact, ultra-low power, intelligent devices for applications in wearables, healthcare, IoT, defense, and wake-up systems, integrating AI into various sensor-based technologies. According to BrainChip, Akida Pico offers the lowest power standalone NPU core (less than 1mW), supports power islands for minimal standby power, and operates within an industry-standard development environment. It’s very small logic die area and configurable data buffer and model parameter memory help optimize the overall die size.
AI era
In the sophisticated artificial intelligence (AI) era of today, including smart technology into consumer items is usually connected with cloud services, complicated infrastructure, and high expenses. Computational power and energy economy are occasionally in conflict in the realm of edge artificial intelligence. Designed for deep learning activities, traditional neural processing units (NPUs) require significant quantities of power, so they are less suited for always-on, ultra-low-power applications including sensor monitoring, keyword detection, and other extreme edge artificial intelligence uses. BrainChip is providing a fresh approach to this challenge.BrainChip’s solution addresses one of the major challenges in edge AI: how to perform continuous AI processing without draining power. Traditional microcontroller-based AI solutions can manage low-power requirements but often lack the processing capability for complex AI tasks.
Steve Brightfield, CMO at BrainChip
2014 saw the launch of BrainChip, which took its inspiration from Peter Van Der Made’s work on neuromorphic computing concepts. Especially using spiking neural networks (SNNs), this technique replicates how the brain manages information, therefore transforming a fundamentally different method to traditional convolutional neural networks (CNNs). The SNN-based systems of BrainChip only compute when triggered by events rather than doing continuous calculations, hence optimizing power efficiency.
In an interview with Embedded, Steve Brightfield, CMO at BrainChip, talked about how this new method will change the game for ultra-low-power AI apps, showing big steps forward in the field. Brightfield said that this new technology makes it possible for common things like drills, hand tools, and other consumer products to have smart features without costing a lot more. “Today, a battery with a built-in tester can show how healthy it is with a simple color code: green means it’s good, red means it needs to be replaced. Providing a similar indicator, AI in these products can tell you when parts are wearing out before they break. BrainChip’s low-power, low-maintenance AI works in the background without being noticed, so advanced tests can be used by anyone without needing to know a lot about them,” Brightfield said.
Traditional NPUs vs. Event-Based Computing
Brightfield claimed that ordinary NPUs—including those with multiplier-accumulator arrays—run on fixed pipelines, processing every input whether or not it is beneficial. Particularly in cases of sparse data, a typical occurrence in AI applications where most input values have little impact on the final outcome, this inefficacy often leads in wasted calculations. By use of an event-based computing architecture, BrainChip saves computational resources and electricity by activating calculations only when relevant data is present.“Most NPUs keep calculating all data values, even for sparse data,” Brightfield remarked. “We schedule computations dynamically using our event-based architecture, so cutting out unnecessary processing.”
The Influence of Sparsity
BrainChip’s main benefit comes from using data and neural weights’ sparsity. Traditional NPU architectures can take advantage of weight sparsity with pre-compilation, benefiting from model weight pruning, but cannot dynamically schedule for data sparsity, they must process all of the inputs.By processing data just when needed, BrainChip’s SNN technology can drastically lower power usage based on the degree of sparsity in the data. BrainChip’s Akida NPU, for instance, could execute only when the sensor detects a significant signal in audio-based edge applications such as gunshot recognition or keyword detection, therefore conserving energy in the lack of acceptable data.
Akida Pico Block Diagram (Source: Brainchip)
Introducing the Akida Pico: Ultra-Low Power NPU for Extreme Edge AI
Designed on a spiking neural network (SNN) architecture, BrainChip’s Akida Pico processor transforms event-based computing. Unlike conventional artificial intelligence models that demand constant processing capability, Akida runs just in response to particular circumstances. For always-on uses like anomaly detection or keyword identification, where power economy is vital, this makes it perfect. The latest innovation from BrainChip is built on the Akida2 event-based computing platform configuration engine, which can execute with power suitable for battery-powered operation of less than a single milliwatt.Wearables, IoT devices, and industrial sensors are among the jobs that call for continual awareness without draining the battery where the Akida Pico is well suited. Operating in the microwatt to milliwatt power range, this NPU is among the most efficient ones available; it surpasses even microcontrollers in several artificial intelligence applications.
For some always-on artificial intelligence uses, “the Akida Pico can be lower power than microcontrollers,” Brightfield said. “Every microamp counts in extreme battery-powered use cases, depending on how long it is intended to perform.”
The Akida Pico can stay always-on without significantly affecting battery life, whereas microcontroller-based AI systems often require duty cycling—turning the CPU off and on in bursts to save power. For edge AI devices that must run constantly while keeping a low power consumption, this benefit becomes very vital.
BrainChip’s MetaTF software flow allows developers to compile and optimize Temporal-Enabled Neural Networks (TENNs) on the Akida Pico. Supporting models created with TensorFlow/Keras and Pytorch, MetaTF eliminates the need to learn a new machine language framework, facilitating rapid AI application development for the Edge.
Akida Pico die area versus process (mm2) (Source: Brainchip)
Standalone Operation Without a Microcontroller
Another remarkable feature of the Akida Pico is its ability to function alone, that is, without a host microcontroller to manage its tasks. Usually beginning, regulating, and halting operations using a microcontroller, the Akida Pico comprises an integrated micro-sequencer managing the full neural network execution on its own. This architecture reduces total system complexity, latency, and power consumption.For applications needing a microcontroller, the Akida Pico is a rather useful co-processor for offloading AI tasks and lowering power requirements. From battery-powered wearables to industrial monitoring tools, this flexibility appeals to a wide range of edge artificial intelligence applications.
Targeting Key Edge AI Applications
The ultra-low power characteristics of the Akida Pico help medical devices that need continuous monitoring—such as glucose sensors or wearable heart rate monitors—benefit.Likewise, good candidates for this technology are speech recognition chores like voice-activated assistants or security systems listening for keywords. Edge artificial intelligence’s toughest obstacle is combining compute requirements with power consumption. In markets where battery life is crucial, the Akida Pico can scale performance while keeping inside limited power budgets.
One of the most notable uses of BrainChip’s artificial intelligence, according to Brightfield, is anomaly detection for motors or other mechanical systems Both costly and power-intensive, traditional methods monitor and diagnose equipment health using cloud-based infrastructure and edge servers. BrainChip embeds artificial intelligence straight within the motor or gadget, therefore flipping this concept on its head.
BrainChip’s ultra-efficient Akida Neural Processor Unit (NPU) for example, may continually examine vibration data from a motor. Should an abnormality, such as an odd vibration, be found, the system sets off a basic alert—akin to turning on an LED. Without internet access or a thorough examination, this “dumb and simple” option warns maintenance staff that the motor needs care instead of depending on distant servers or sophisticated diagnosis sites.
“In the field, a maintenance technician could only glance at the motor. Brightfield said, “they know it’s time to replace the motor before it fails if they spot a red light.” This method eliminates the need for costly software upgrades or cloud access, therefore benefiting equipment in distant areas where connectivity may be restricted.
Regarding keyword detection, BrainChip has included artificial intelligence right into the device. According to Brightfield, with 4-5% more accuracy than historical methods using raw audio data and modern algorithms, the Akida Pico uses just under 2 milliwatts of power to provide amazing results. Temporal Event-Based Neural Networks (TENNS), a novel architecture built from state space models that permits high-quality performance without the requirement for power-hungry microcontrollers, enable this achievement.
As demand for edge AI grows, BrainChip’s advancements in neuromorphic computing and event-based processing are poised to contribute significantly to the development of ultra-efficient, always-on AI systems, providing flexible solutions for various applications.
Tags:
Maurizio Di Paolo Emilio
Last edited: