Sumeet Kumar, Co-founder and CEO of Innatera – Interview Series

Sumeet Kumar is the Co-founder and CEO of Innatera Nanosystems, where he leads the development of ultra-efficient neuromorphic processors for sensor data analytics in IoT, wearable, and embedded devices. Previously, he managed EU-funded research at TU Delft, including the €50 million PRYSTINE project on autonomous vehicle technologies, and held roles at Intel and in academic research on advanced processor architectures.

Innatera is a semiconductor company focused on bringing ultra-low-power intelligence to the “sensor edge.” Their core innovation lies in spiking neural processors built on an analog–mixed signal architecture that mimic the brain’s event-based processing. These chips can recognize patterns in sensor data at sub-milliwatt power levels and ultra-low latency, making them ideal for always-on, power-constrained applications.

You co-founded Innatera in 2018 with a vision to bring neuromorphic processors from the lab into real-world devices. What drove you personally to start the company, and how has that vision evolved over the past seven years?

Innatera was founded with a clear mission in mind: to bring brain-like intelligence directly to the sensor. The signs were clear even back in 2018, devices were integrating increasingly complex sensors, and the need for always-on sensing was growing. Microcontrollers lacked power-efficient AI capabilities, and even that would only move the needle so far when it came to continuous processing in devices powered by small batteries. It was clear that the way sensor data is processed in these devices needed to change, and the decade of research we had done at the TUDelft on neuromorphic computing and energy-efficient processing seemed to have an answer to this challenge.

Our vision has remained consistent – a smarter, cleaner and safer world powered by ambient intelligence. By bringing intelligence to sensors, our chips will enable the world’s sensor data to be processed directly at the source, leading to a radical reduction in the energy use of modern AI. We aim to make a billion sensors intelligent by 2030.

Pulsar is the first step in that journey – it is the world’s first neuromorphic microcontroller designed for mainstream adoption. It makes brain-inspired intelligence practical in wearables, smart home devices, and industrial systems, among other use cases, while laying the foundation for adaptive, autonomous technologies of the future.

Pulsar is based on a fundamentally new approach to processing at the sensor, realized over 7 years of hard research and engineering. What began as a venture with four people, has in this time grown to a global team of 100, across 15 countries, united by a people-first culture built on resilience, creativity, and ambition.

Pulsar is described as the first truly mass-market neuromorphic microcontroller. What makes it different from previous neuromorphic chips that remained largely confined to research labs?

The focus of academic research is often on developing innovative new approaches to solve tough problems. As a consequence, the benefits of solutions tend to be measured in isolation. However, when these new technologies are deployed into production, they have to interact with other parts of the system, which often results in their benefits being diluted. This is also the case for a lot of neuromorphic and conventional AI acceleration technologies – they are integrated into systems that aren’t designed with the same fundamentals in mind, and this leads to an outcome that pales in its efficiency. Pulsar, on the other hand, is a complete, standalone microcontroller, purpose built for efficient processing of sensor data at the extreme edge.

It was designed from the ground up to integrate everything needed for sensor data processing within a single chip: analog and digital spiking neural cores, CNN and FFT accelerators, and a full 32-bit RISC-V subsystem for system management and sensor control. This heterogeneous architecture allows Pulsar to transform raw sensor data into actionable insights directly at the device, while consuming up to 500x less energy and running 100x faster than conventional AI processors.

Beyond hardware, Pulsar also addresses the long-standing software barrier. Its Talamo SDK, with native PyTorch integration, makes neuromorphic development accessible to mainstream engineers and enables compact models under 5KB to run in sub-milliwatt power budgets. Fitting all of this into a 2.8 x 2.6 mm package, Pulsar eliminates the need for bulky multi-chip setups, making it the first neuromorphic processor ready for true mass-market deployment.

Accessibility is a big theme for Innatera. How does the Talamo SDK, particularly with its PyTorch integration, lower the barrier for developers who are new to neuromorphic computing? 

For decades, the main barrier to neuromorphic adoption wasn’t due to the hardware, but instead was because of a lack of developer-friendly tools. Developers were faced with steep learning curves and unfamiliar workflows, which in turn slowed innovation. Talamo directly addresses this by providing a PyTorch-based SDK that lets engineers design, train, and deploy spiking neural networks through familiar workflows. Compact models can be easily integrated into existing sensor architectures, enabling always-on intelligence in even the smallest, most power-constrained devices. By removing complexity and speeding development, Talamo makes neuromorphic computing accessible to mainstream developers and accelerates the path from prototype to product.

From a technical standpoint, how do you balance the analog and digital spiking accelerators inside Pulsar to handle diverse workloads efficiently? 

Pulsar’s architecture blends analog and digital spiking cores to optimize energy use and flexibility. The analog cores provide ultra-efficient processing for continuous, always-on sensor workloads where every microwatt counts. The digital cores deliver programmability and precision for more complex or variable tasks, still within an efficient power envelope. Workloads are distributed across the two depending on application needs, ensuring that energy is only consumed when data changes. This event-driven approach allows Pulsar to sustain sub-milliwatt performance while maintaining the flexibility to support diverse real-world applications.

Can you walk us through a typical developer workflow—from training a model to deploying it on Pulsar—and where the biggest efficiency gains are realized?

The workflow begins in PyTorch, where developers design and train their models as they would for conventional AI. Using Innatera’s Talamo SDK, the model is converted into a spiking neural network optimized for Pulsar’s hardware. Developers can then simulate, refine, and deploy the model directly onto the chip, often with footprints as small as 5KB. The model development step is integrated within a larger application pipeline development flow that enables the developer to build code that targets the RISC-V, as well as the CNN accelerator, in a unified manner. This translates to an improved development experience, and shorter development time.

The biggest efficiency gains occur once the model is up and running on Pulsar’s event-driven spiking cores. Unlike conventional MCUs that burn power continuously, Pulsar computes only when input data changes. This allows always-on tasks such as gesture recognition or radar presence detection to run continuously at sub-milliwatt levels, delivering orders-of-magnitude improvements in energy efficiency while maintaining high accuracy, and incredibly short latency.

Which sectors are showing the fastest adoption of your technology, and can you share examples of early customers or partners who are already deploying Pulsar in products?

Adoption of Pulsar is happening fastest in areas where always-on sensing and ultra-low power matter most, including smart homes, wearables, and industrial safety. A good example of this is Aaroh Labs, which has developed next-generation smoke detectors powered by Innatera, recently unveiled at SEMICON India 2025. These devices do more than just sense smoke by combining smoke detection with human-presence monitoring, creating richer situational awareness and enabling smarter safety systems for residential, commercial, and industrial environments.

The same neuromorphic approach can extend to asset tracking and environmental monitoring, with broad implications for connected healthcare and smart cities. At SEMICON India, CYRAN AI Solutions also showcased how Innatera’s technology is being integrated into compact sensor systems such as electromyography (EMG) wearables for gesture recognition, highlighting the potential of neuromorphic AI to enable intuitive human-machine interaction.

These early deployments are only the beginning, signaling that neuromorphic computing is shifting from theory to practice as we speak, and rapidly taking root in real-world applications.

In demonstrations we’ve seen examples like ultra-low-power gesture recognition and radar-based presence detection consuming under a milliwatt. How do you validate accuracy and reliability in such constrained environments?

Validation often depends on the application, and in addition to accuracy, false positive and false negative detection rates provide a critical indication of the reliability of a solution. Often, customers have specific KPIs and test conditions for validation. Pulsar’s flexibility is key in enabling comprehensive solutions that enable customers to check all the boxes for their use-case. Comparisons are done by benchmarking against conventional MCUs and accelerators, which typically consume 40-100 times more power for the same tasks.

In real-world demonstrations, such as radar-based presence detection and audio scene classification, Pulsar consistently delivers accuracies above 90% while staying within sub-milliwatt budgets. This enables continuous operation without sacrificing reliability, something traditional always-on systems traditionally had to compromise on by waking from sleep, throttling performance, or offloading to the cloud.

You’ve positioned Pulsar as complementary to more conventional NPUs and CPUs. How do you see neuromorphic computing fitting into the broader silicon stack of future smart devices?

Pulsar is designed as the first chip that sensors communicate with. It processes data locally at ultra-low power, converting raw sensor signals into meaningful, actionable information, directly at the source. NPUs and CPUs can then be engaged only when heavier processing is required.

This makes neuromorphic processors a complementary layer in the silicon stack; an always-aware, always-on foundation that extends device lifespans, reduces energy use, and improves responsiveness. Pulsar takes the task of sensor data processing away from the traditionally higher-power components in the system, allowing them to be switched off in many devices, and in some cases, even eliminated completely. This leads to smarter and longer lasting devices.

What role do collaborations with partners like Aaroh Labs and CYRAN AI Solutions play in accelerating real-world adoption of neuromorphic AI?

Partnerships act as the bridge between breakthrough technology and broad adoption. By working with innovators like Aaroh Labs and CYRAN AI Solutions, Innatera ensures Pulsar is validated in real-world environments and tailored for specific verticals. Aaroh Labs brings neuromorphic intelligence into critical safety infrastructure, while CYRAN AI Solutions demonstrates its potential in intuitive human-machine interaction. These collaborations prove the technology’s versatility, reducing barriers for other adopters and building confidence in deploying neuromorphic processors at scale.

Our partnerships with sensor vendors such as Socionext allow us to closely pack intelligence into the sensor module, simplifying the uptake and deployment of intelligent sensing into devices. Further, such collaborations strengthen our already strong and growing ecosystem, and accelerate the spread of neuromorphic computing in the industry.

Looking ahead, do you see Pulsar and its successors moving toward on-device learning and adaptation, rather than just inference at the edge?

Absolutely. With Pulsar, we’ve only just begun to scratch the surface of what neuromorphic can do. Neuromorphic processors are inherently well-suited to online learning and adaptation, and Pulsar lays the groundwork for devices that can do far more than just detecting and responding.

Neuromorphic computing is set to enable a new generation of adaptive and autonomous edge devices; systems that learn, self-calibrate, and optimize in real time while running on tiny batteries. This evolution will unlock a wide range of applications, ranging from wearables that adjust to your behavior on the fly, to industrial systems that predict and prevent failures with minimal energy use. The long-term goal is to create devices that are just as intelligent as they are continuously adaptive and resilient, redefining what’s really possible at the edge.

Thank you for the great interview, readers who wish to learn more should visit Innatera.