Periodic Reporting for period 2 - Spiking Neural Processor (Always-On Intelligent Sensing with the Spiking Neural Processor)
Okres sprawozdawczy: 2024-05-01 do 2025-04-30
Innatera aims to bring next-generation AI capabilities directly to every sensor in the world. It does this through brain-inspired microprocessors that enable advanced signal processing and pattern recognition capabilities, within an ultra-narrow power envelope. By enabling intelligent sensing, Innatera’s microprocessors unlock the true potential of sensors in portable and wearable devices that run on small batteries. Emerging application use-cases for sensors tend to be in small consumer devices, which are tremendously constrained in their (1) power dissipation, (2) size, and (3) cost, while under continuous market pressure to introduce new application features. Meeting these often contradictory requirements necessitates more powerful AI models, running on more compact and more energy efficient processing hardware. The EIC Spiking Neural Processor project aims to prototype a radical new architecture for near-sensor processing with a novel analog-mixed signal computing fabric. The project will deliver a silicon prototype that improves application performance, power dissipation, and integration density, while concomitantly realizing disruptive new functional capabilities for sensor systems.
The project’s outcomes will enable sensor systems to intelligently make sense of the data they produce, and eliminate the need to send all captured data into the cloud for processing. This will allow massive energy savings, since over 95% of captured data in common application use-cases actually contains no information of relevance. Applied to the billions of sensors sold every year, the developed technologies hold the potential to achieve CO2 reductions in the millions of tonnes annually.
The project’s outcomes will enable sensor systems to intelligently make sense of the data they produce, and eliminate the need to send all captured data into the cloud for processing. This will allow massive energy savings, since over 95% of captured data in common application use-cases actually contains no information of relevance. Applied to the billions of sensors sold every year, the developed technologies hold the potential to achieve CO2 reductions in the millions of tonnes annually.
The EIC Spiking Neural Processor was successfully executed over a period of two years between May 2023-April 2025. The project aimed to assess market requirements to provide a technical direction for R&D activities, develop new IP to enable novel functional capabilities and energy efficiency gains, and integrate these within a silicon prototype for validation. These activities were successfully carried out within the project period.
In year one of the project, activities focussed on four key topics:
1. Improving the performance of processing architectures for spiking neural network inference across application use-cases
2. Optimizing the computational fabric of the Spiking Neural Processor for power dissipation, robustness, and manufacturability
3. Development of interfaces to targeted sensor systems relevant to beach-head markets
4. Extension of software support targeting the newly developed IP
Year one commenced with an intensive requirements gathering exercise through which the beach-head market and application use-cases were identified, in close consultation with early customers. Application requirements were defined for each targeted sensor, and this was used as guidance for the subsequent design space exploration. Architecture exploration activities began on the four topics listed above, with the intention of identifying an optimal technical solution for implementation. Simulation frameworks were developed and the first proof-of-concept validations for the new IP blocks were realized in year one. Subsequently, hardware implementations were also realized.
Activities in year two of the project focussed on 4 key topics:
1. Design and backend implementation of the improved computational fabric and interface IP
2. Implementation of an improved architecture to integrate the newly developed IP
3. Silicon prototyping of the new blocks and characterization
4. Application software development and mapping to the new prototype
Year two involved realizing the new architectures explored in the first year of the project, resulting in the creation of a silicon prototype that enables characterization of the new IP. Using the software tooling developed in the first year, a pilot application was mapped onto the developed IP for testing. The outcome at the end of year two is a commercially exploitable architecture that will be applied to Innatera’s next product to be released in 2026-2027.
In year one of the project, activities focussed on four key topics:
1. Improving the performance of processing architectures for spiking neural network inference across application use-cases
2. Optimizing the computational fabric of the Spiking Neural Processor for power dissipation, robustness, and manufacturability
3. Development of interfaces to targeted sensor systems relevant to beach-head markets
4. Extension of software support targeting the newly developed IP
Year one commenced with an intensive requirements gathering exercise through which the beach-head market and application use-cases were identified, in close consultation with early customers. Application requirements were defined for each targeted sensor, and this was used as guidance for the subsequent design space exploration. Architecture exploration activities began on the four topics listed above, with the intention of identifying an optimal technical solution for implementation. Simulation frameworks were developed and the first proof-of-concept validations for the new IP blocks were realized in year one. Subsequently, hardware implementations were also realized.
Activities in year two of the project focussed on 4 key topics:
1. Design and backend implementation of the improved computational fabric and interface IP
2. Implementation of an improved architecture to integrate the newly developed IP
3. Silicon prototyping of the new blocks and characterization
4. Application software development and mapping to the new prototype
Year two involved realizing the new architectures explored in the first year of the project, resulting in the creation of a silicon prototype that enables characterization of the new IP. Using the software tooling developed in the first year, a pilot application was mapped onto the developed IP for testing. The outcome at the end of year two is a commercially exploitable architecture that will be applied to Innatera’s next product to be released in 2026-2027.
The activities of the project resulted in the following key outcomes:
1. A new set of interfaces for smart sensing applications
2. Improved SNN acceleration fabric for use in applications requiring longer-time scales
3. Extended software tooling to leverage the new functional capabilities.
4. New prototype applications - e.g. gesture recognition with IMUs
5. Commercial feedback from target customers validating the key value propositions of the SNP-II
architecture, and its competitive differentiation.
These outcomes will be commercialized as part of new production introductions, and thus intercept billions of new sensors in the period between 2026 and 2030.
1. A new set of interfaces for smart sensing applications
2. Improved SNN acceleration fabric for use in applications requiring longer-time scales
3. Extended software tooling to leverage the new functional capabilities.
4. New prototype applications - e.g. gesture recognition with IMUs
5. Commercial feedback from target customers validating the key value propositions of the SNP-II
architecture, and its competitive differentiation.
These outcomes will be commercialized as part of new production introductions, and thus intercept billions of new sensors in the period between 2026 and 2030.