Introducing SpiNNaker2 – The Future of Hybrid Brain Inspired High Performance Computing

Dresden, May 8, 2024

Hector Gonzalez, Co-CEO and Co-Founder

In 1985, Steve Furber introduced the ARM core to the world, a deeply disruptive innovation that now powers over 260 billion mobile devices globally. At the time, the technological and commercial potentials of this development were unimaginable even to its creators. But this was just the beginning of Furber’s impact on computing technology.

Nearly 35 years later, in 2019, Furber embarked on another ambitious project, when the SpiNNaker1 machine was launched at the University of Manchester. The development of the SpiNNaker1 technology was part of the €1 billion EU-funded Human Brain Project and represented a significant leap in computing architecture. SpiNNaker1 was designed as a massively parallel, manycore computer system aiming to emulate brain models. This development not only advanced our understanding of cognitive functions, it also paved the way for new potential applications in computing just the way the Arm core eventually did.

The Advent of SpiNNaker2 as an event-based platform for Hybrid AI

Fast forward to 2024, where we at SpiNNcloud Systems in Dresden, Germany, today publicly introduce the next iteration of this technology. Developed predominantly at the chair of Prof. Dr. Christian Mayr at the Technical University Dresden, the SpiNNaker2 system has evolved from its predecessor to become an event-based platform for AI while enhancing significantly all its original features.

The technology we are presenting today offers a path to address inherent limitations faced by traditional AI models. Current models, while powerful and ever so more omnipresent in all of our lives, suffer from transparency issues, lack of explainability, require vast amounts of training data, and as Steve Furber mentions in the video, they burn a ridiculous amount of power. These issues stem from the limitations of the three main types of AI:

1) Deep Neural Networks (also known as connectionists or statistical approaches): They have shown amazing practical applications by properly handling very large datasets and complex computations. However, they are often not explainable, nor do they offer a generalizable alternative to extrapolate knowledge from different contexts and their size also makes them extremely inefficient.

2) Symbolic AI/Expert Systems: They offer excellent explainability and operate under strict rulesets, so they are highly reliable, traceable, and explainable. However, they rely on experts to encode their knowledge, and adapt poorly to new situations as they are often not exposed to the huge datasets in the statistical approaches.

3) Neuromorphic Models: These brain-inspired neural networks propel some of the most intriguing principles of the human brain, for instance the energy efficiency, the event-based operation, and the highly parallel approach. However, they tackle the brain modeling process through a detailed bottom-up approach that often leads to the oversimplification of cognitive processes.

To address these challenges, hybrid AI models combine the strengths of these systems to enhance robustness, scalability, practicality, and energy efficiency. This advanced combination of different AI principles is often referred to as “The Third Wave of AI”, a term coined by DARPA, which aims to describe systems that can understand contextual nuances and adapt dynamically. The heterogeneity of these models has an extremely high chance of eventually bridging the gap between human-like reasoning and machine efficiency, paving the way towards a universally usable AI with an acceptable margin of error.

Now what does this have to do with hardware. and, more specifically, with SpiNNaker2? Well, it turns out that GPUs and the new wave of industrial approaches to accelerate Deep Neural Networks have strongly focused in designing efficient machinery to compute the numeric required in these networks without leaving room for a rule-based programmability that is entangled with these accelerators. Furthermore, this new wave of accelerators is still designed under the synchronous primitives inherited from GPUs, which prevents them from aiming to be efficient like our brains. Hence, a hardware platform enabling rule-based engines highly coupled with machinery to accelerate deep neural networks, plus operating efficiently like our brains, have the potential to revolutionize the large-scale implementation of these robust models. One particular algorithm example, NARS-GPT, which has outperformed GPT-4 at reasoning tasks, uses a neuro-symbolic engine that is almost impossible to parallelize in an HPC system based on GPUs or any other dataflow-driven hardware. NARS, short for Non-Axiomatic Reasoning System, uses Deep Neural Networks to extract features from the environment, which are further passed to a knowledge graph created with a rule-based (or symbolic) approach. To the best of our knowledge, and the knowledge of the authors of these neuro-symbolic reasoners, hybrid hardware is the only possibility to scale up these models and deploy them efficiently. These features offer the potential to integrate learning, reasoning, and efficient processing not only at a chip- but also at a system-level.

Building the Future of HPC with SpiNNaker2

As teased earlier, the SpiNNaker2, main building block of our solution, is specifically serving these sophisticated demands. Each chip is a low-power mesh of 152 ARM-based cores, which enhances parallel processing capabilities crucial for managing the complex, dynamic data typical in hybrid AI systems. Its architecture supports energy-proportional high-speed interconnects, which are essential for scaling and communicating across multiple chips, bolstering its capacity for large-scale, distributed neural network deployment. The chip also incorporates native accelerators specifically designed for neuromorphic computing tasks such as exponential, and logarithmic functions as well as true random number generation, alongside custom accelerators for efficient machine learning computations.

This scalability is facilitated by the chip’s lightweight and event-based Network-on-Chip architecture, which supports rapid and efficient data transfer across the system, crucial for tasks that require real-time processing across many nodes, such as complex AI computations and large-scale neural simulations. Additionally, the globally asynchronous and locally synchronous (GALS) architecture allows each portion of the chip to operate in an interrupt-driven approach, which significantly reduces bottlenecks and enhances the overall system performance while scaling up. Furthermore, the system’s native support for dynamic voltage frequency scaling at the core level contributes to a scalable energy management approach that adjusts power in an on-demand basis. This feature contributes to maintaining the efficiency of the system at all levels of abstraction.

Now as a Neuromorphic supercomputer, the SpiNNcloud platform remains an extremely competitive platform reaching scalability levels (i.e., number of neurons) that are not possible today for any other system in the planet. The SpiNNcloud platform made available in Dresden has the capacity to emulate at least 5 billion neurons. Additionally, the SpiNNaker2 architecture highlights through its flexibility, allowing the native implementation of not only Deep Neural Networks, Symbolic models, or Spiking Neural Networks, but pretty much any other computation that can be represented in a computational graph. With such extensive scalability, the SpiNNcloud platform is engineered to achieve Supercomputer Performance Levels while maintaining high peak efficiency and reliability. This makes it suitable for handling sophisticated AI challenges like the ones that lie before us in the Third Wave of AI.

Steve Furber once mentioned to me in an interview that while the Arm technology had a significant impact, he sees the SpiNNaker invention as a more disruptive and fundamentally profound contribution because it aims to unveil practical inspiration from the mysteries lying behind the human brain.