For Really Low Power Neural Networks Try Spiking Neural Networks
20
Companies
46
Key Quotes
3
Topics
1
Insights
🎯 Summary
Inside Electronics Podcast Summary: Pulsar and the Future of Edge AI with Spiking Neural Networks
This episode of the Inside Electronics Podcast, hosted by Bill Long and featuring Samit (CEO and co-founder of Inaterra), centers on the commercialization of Spiking Neural Networks (SNNs) through their novel chip implementation called Pulsar. The core narrative arc moves from the limitations of traditional microcontrollers in handling modern sensor data to the resource-efficient solution offered by neuromorphic computing.
Key Takeaways for Technology Professionals
1. The SNN Paradigm Shift for Edge AI
- SNN vs. Traditional NN: Traditional Neural Networks (NNs) are computationally intensive due to continuous data abstraction and high data movement. SNNs encode information in time (sparse spikes of 0 or 1) rather than magnitude (width), leading to models that are approximately 100 times smaller than equivalent Deep Neural Networks (DNNs) for the same accuracy.
- Inherent Temporal Processing: SNNs naturally understand time sequences, making them ideal for processing continuous, time-series sensor data without complex topologies.
2. Pulsar: A Neuromorphic Microcontroller Solution
- Addressing the MCU Gap: Traditional MCUs were struggling to keep up with complex sensor data streams, forcing AI burdens onto larger MPSoCs or the cloud, increasing power and latency. Pulsar fills this gap by offering powerful AI directly on the microcontroller level.
- Mixed-Signal Architecture: Pulsar employs a mixed-signal acceleration approach to balance performance across different time scales:
- Analog Fabric: Used for high energy efficiency and excellent temporal resolution (nanosecond to microsecond events).
- Digital Fabric: Used for flexibility and handling longer time scales (hundreds of milliseconds to seconds), albeit with slightly higher energy consumption than the analog core.
- Integrated Compute: Pulsar is designed as an all-in-one component, integrating the SNN accelerator, a CNN accelerator, an FFT accelerator, and a 32-bit integer floating-point RISC-V core for system management and command execution. This integration minimizes the Bill of Materials (BOM) for edge devices.
3. Strategic Business Implications and Target Applications
- Breaking the Trade-off Triangle: Pulsar aims to eliminate the traditional trade-off between functionality, accuracy, and power consumption at the edge.
- Target Use Cases: The primary focus is on sensor-heavy applications requiring continuous, low-latency, high-accuracy processing on limited power budgets (e.g., battery-operated devices).
- Presence Sensing (Radar): A reference design with Solonix demonstrated extending battery life in smart doorbells from 3-4 weeks to 1.5 years by using the SNN to continuously monitor radar data and only waking up the power-hungry camera upon human detection.
- People Counting (Infrared): Working with Melexis, SNNs differentiate human heat signatures from environmental noise (like coffee cups) in battery-powered smoke detectors for emergency response.
- Data Privacy: Keeping powerful AI processing local to the sensor inherently supports data privacy requirements.
4. Technical Implementation and Tooling
- Encoder/Decoder Efficiency: The energy cost of converting conventional data into the spike-based representation (encoders) and back (decoders) is negligible compared to the massive power savings of the SNN core itself, avoiding the common pitfall of analog computing.
- Software Development Kit (Tulumo): Inaterra provides the Tulumo SDK to simplify deployment. Crucially, the SNN model building component is integrated with PyTorch via an extension, allowing developers to use familiar workflows before the toolchain maps the model onto the Pulsar hardware architecture seamlessly.
- Security: The initial version relies primarily on software-level security, with hardware security extensions planned for future derivatives targeting specific regulated markets.
5. Future Outlook
The team believes they are only “scratching the surface” of SNN potential, citing complex biological processes like noise cancellation in crowded rooms as future capabilities that could be unlocked by further exploiting the temporal dynamics of neuromorphic hardware.
🏢 Companies Mentioned
Venetian
✅
hospitality
Las Vegas
✅
unknown
And I
✅
unknown
Delft University
✅
unknown
Inside Electronics
✅
unknown
Senior Content Director
✅
unknown
Bill Long
✅
unknown
Alex Palt
✅
unknown
Electronic Design
✅
unknown
Inside Electronics Podcast
✅
unknown
Melexis
🔥
tech
Solonix
🔥
tech
RISC-V
🔥
tech
Tulumo
🔥
tech
PyTorch
🔥
tech/Software
💬 Key Insights
"we consume such a small amount of energy that with our solution plugged into a standard doorbell, you extend battery life from three to four weeks between recharges to one and a half years between recharges, which is a staggering improvement in battery life without affecting any of the coverage."
"Effectively, what that translates to is that it makes Pulsar into an all-in-one component. You literally have everything that you need to go from raw sensor data to an actionable insight in just one place."
"all batteries, are located in really small housings where you cannot burn a lot of power, where user experience requires that latency be kept really short, accuracy be high, data be kept private to the device. And all of that means that you've got to have really powerful AI sitting right next to the sensor."
"With traditional AI, you often run into this trade-off between functionality, accuracy, and power. But with Pulsar, you don't really have to do that anymore."
"Spiking models tend to be about a hundred times smaller than their equivalent deep neural network counterparts at the same level of accuracy for the same applications."
"We knew that the brain-inspired spiking neural networks tend to be a lot more powerful than traditionally AI approaches. And they can do a lot more with models that are about a hundred times smaller."
📊 Topics
#artificialintelligence
45
#startup
1
#aiinfrastructure
1