Periodic Reporting for period 1 - SPINNODE (SpiNNaker on the Edge)
Berichtszeitraum: 2023-05-01 bis 2024-04-30
SpiNNcloud is a deep-tech company commercializing brain-inspired systems for Artificial Intelligence workloads, optimisation tasks, drug discovery processes, and more. The technological solution offered is scalable from a single chip level up to supercomputer level comprising 69.120 SpiNNaker2 chips at the largest server size and thus serves use cases at both the edge and the cloud. The technology is mature, and the company has initial success engaging with different well-known institutions.
To enable early access to the technology and engage with a wider number of stakeholders, the SpiNNaker2 technology is also brought to the edge. The main objective of the SpiNNode project is to further increase the maturity of SpiNNcloud’s technology through easy-to-use software and to engage with different stakeholders through granting access to a single-chip PCB, created during the project. The SpiNNode project also aims to materialize a detailed business plan for further commercialization of the technology.
Precise goals along the execution of the project are:
Develop a multi-processing-element deployment framework for distributing deep neural networks (DNN) into SpiNNaker2-based systems at the edge and the cloud level.
Develop an optimised toolchain for spiking neural networks (SNN) including partitioning, mapping, and deployment.
Develop the very first multi-core hybrid framework that allows the combined execution of DNNs and SNNs into the same hardware.
Create a brain-inspired and energy efficient edge PCB system called SpiNNode board based on the SpiNNaker2 ASIC.
Deploy the SpiNNode board in a variety of applications to validate its performance and showcase its operation.
Disseminate the project results by engaging with a wide range of stakeholders, ensuring that the IP is protected
Prepare the commercial uptake of the innovation to exploit the results obtained in the project.
To enable early access to the technology and engage with a wider number of stakeholders, the SpiNNaker2 technology is also brought to the edge. The main objective of the SpiNNode project is to further increase the maturity of SpiNNcloud’s technology through easy-to-use software and to engage with different stakeholders through granting access to a single-chip PCB, created during the project. The SpiNNode project also aims to materialize a detailed business plan for further commercialization of the technology.
Precise goals along the execution of the project are:
Develop a multi-processing-element deployment framework for distributing deep neural networks (DNN) into SpiNNaker2-based systems at the edge and the cloud level.
Develop an optimised toolchain for spiking neural networks (SNN) including partitioning, mapping, and deployment.
Develop the very first multi-core hybrid framework that allows the combined execution of DNNs and SNNs into the same hardware.
Create a brain-inspired and energy efficient edge PCB system called SpiNNode board based on the SpiNNaker2 ASIC.
Deploy the SpiNNode board in a variety of applications to validate its performance and showcase its operation.
Disseminate the project results by engaging with a wide range of stakeholders, ensuring that the IP is protected
Prepare the commercial uptake of the innovation to exploit the results obtained in the project.
The SpiNNode project has achieved a software framework to partition and map Deep Neural Networks into SpiNNaker2-based systems. It has enabled the creation of the low-level firmware to bring-up test and boot complex multi-chip systems such as the 48-node board, as well as set the foundation for the bare-metal operating system managing the resources in SpiNNaker2-based systems. Such minimal operating system is based on three components: SARK2 (SpiNNaker Application Runtime Kernel): a minimal set of libraries and functions to use the resources in the chip during operation; SCAMP2: a monitor core in charge of bringing-up the system and enabling communications within the system; and Spin2-API: an event-based operating system that allows the execution of interrupt-driven software across the numerous processing elements.
Along the Spiking Neural Networks field, the SpiNNode project has also made possible the creation of a two-repositories project with high-level python bindings to enable easy-to-use execution of models into SpiNNaker2-based hardware without the complexity of mapping directly bare-metal models. Using these software repositories, it has been possible to host events in workshops such as Capocaccia and Telluride in which external teams have interacted with our technology. As a result of this, publications along the lines of mapping QUBO-based optimisation problems, comparing SpiNNaker2 within the Neuromorphic systems at scale in Nature communications, and engaged with the community to be part of Neuromorphic benchmark initiatives such as Neurobench.
Additionally, the SpiNNode project has also derived the specifications and design of the single-chip PCB through numerous interactions with stakeholders that span from event-based camera providers, manufacturers of underwater robots, humanoid robots, to electronic providers for drone systems. During the SpiNNode project, a detailed characterisation of the SpiNNaker2 thermal behaviour was accomplished, as well as the design of a mechanical, electrical and thermal concept for the board. The prototype board is currently in production.
Along the Spiking Neural Networks field, the SpiNNode project has also made possible the creation of a two-repositories project with high-level python bindings to enable easy-to-use execution of models into SpiNNaker2-based hardware without the complexity of mapping directly bare-metal models. Using these software repositories, it has been possible to host events in workshops such as Capocaccia and Telluride in which external teams have interacted with our technology. As a result of this, publications along the lines of mapping QUBO-based optimisation problems, comparing SpiNNaker2 within the Neuromorphic systems at scale in Nature communications, and engaged with the community to be part of Neuromorphic benchmark initiatives such as Neurobench.
Additionally, the SpiNNode project has also derived the specifications and design of the single-chip PCB through numerous interactions with stakeholders that span from event-based camera providers, manufacturers of underwater robots, humanoid robots, to electronic providers for drone systems. During the SpiNNode project, a detailed characterisation of the SpiNNaker2 thermal behaviour was accomplished, as well as the design of a mechanical, electrical and thermal concept for the board. The prototype board is currently in production.
One important aspect to highlight is that the software developments carried out through the EIC transition project are not applications themselves. These are pure frameworks to enable easy utilization of SpiNNaker2 technology, which is a crucial milestone for achieving technological market fit. Hence, what has materialized as result of the project is the creation of a novel multi-processing-element framework that can partition DNN- and SNN-based models into Globally Asynchronous Locally Synchronous (GALS) architectures. The concepts and developments achieved then find applications not only in SpiNNaker2-based systems but also in mesh-based architectures with several constrained nodes. One example of this is systems on chip with several processing elements and accelerators.
Besides the framework itself, research carried out in the SpiNNode project has also enabled QUBO-based solvers for optimisation problems, as well as highly parallel drug discovery based on smaller models that are deployed across the numerous processing elements within the SpiNNaker2 fabric. Despite these two approaches, not being the focus of the SpiNNode project, the frameworks developed by the project have made them possible, enabling a potential exploitation of these along SpiNNcloud’s go to market strategy.
Besides the framework itself, research carried out in the SpiNNode project has also enabled QUBO-based solvers for optimisation problems, as well as highly parallel drug discovery based on smaller models that are deployed across the numerous processing elements within the SpiNNaker2 fabric. Despite these two approaches, not being the focus of the SpiNNode project, the frameworks developed by the project have made them possible, enabling a potential exploitation of these along SpiNNcloud’s go to market strategy.