Source-linked AI summary
SpiNNaker 2: A 10 Million Core Processor System for Brain Simulation and Machine Learning
Christian Mayr, Sebastian Hoeppner, Steve Furber
TL;DR
The paper addresses how to scale SpiNNaker’s real-time spiking-neural-network simulation while improving power efficiency and computational breadth. It proposes SpiNNaker2, combining tenfold core scaling with adaptive power management and numerical accelerators. The design targets approximately 50-fold greater simulation capacity and applications spanning brain simulation, machine learning, and real-time interaction.
Problem
SpiNNaker must scale spiking-neural-network simulation beyond the first-generation 1 Million-core system while preserving energy-efficient real-time operation.
Method
SpiNNaker2 combines processor-based flexibility with dynamic voltage and frequency scaling, power-aware communication, and multiply-accumulate arrays.
Results
Approximately 50× greater overall simulation capacity is expected for SpiNNaker2 than the first-generation system.
Takeaways & Limitations
SpiNNaker2 is intended to support larger and more complex spiking models while extending use toward machine learning and real-time mobile human-machine interaction.
Abstract
from arXiv · showhide
SpiNNaker is an ARM-based processor platform optimized for the simulation of spiking neural networks. This brief describes the roadmap in going from the current SPINNaker1 system, a 1 Million core machine in 130nm CMOS, to SpiNNaker2, a 10 Million core machine in 22nm FDSOI. Apart from pure scaling, we will take advantage of specific technology features, such as runtime adaptive body biasing, to deliver cutting-edge power consumption. Power management of the cores allows a wide range of workload adaptivity, i.e. processor power scales with the complexity and activity of the spiking network. Additional numerical accelerators will enhance the utility of SpiNNaker2 for simulation of spiking neural networks as well as for executing conventional deep neural networks. These measures should increase the simulation capacity of the machine by a factor $>$50. The interplay between the two domains, i.e. spiking and rate based, will provide an interesting field for algorithm exploration on SpiNNaker2. Apart from the platforms' traditional usage as a neuroscience exploration tool, the extended functionality opens up new application areas such as automotive AI, tactile internet, industry 4.0 and biomedical processing.
Introduction
SpiNNaker is an ARM-based platform designed for energy-efficient real-time simulation of spiking neural networks. SpiNNaker2 targets tenfold core scaling within the same power budget and approximately 50-fold greater simulation capacity.
- Architecture: SpiNNaker integrates many ARM cores with communication and memory access optimized for neural-network simulation.Each core independently simulates neurons and exchanges activity through a lightweight asynchronous spike protocol.
- SpiNNaker1 baseline: 1 Million ARM processors make the first-generation system capable of simulating spiking neural networks on the order of 1% of the human brain.
- SpiNNaker2 targets: 10× more cores are targeted for SpiNNaker2 while remaining within the same power budget, corresponding to 10× better power efficiency.
- SpiNNaker2 targets: Approximately 50× greater overall simulation capacity is expected from SpiNNaker2.
1. Spinnaker2 capabilities and main new building blocks
SpiNNaker2 preserves processor-based flexibility while adding workload-adaptive power management, streamlined communication, and numerical acceleration. These building blocks are intended to improve efficiency across changing computational loads.
- Design approach: SpiNNaker2 retains processor-based flexibility while adding specialized support rather than performing every operation in software.
- Adaptive power: Dynamic voltage and frequency scaling adjusts each processor to incoming spike load and required computation every millisecond.
- Adaptive power: Lower computational loads can be stretched over time and run at lower supply voltage, reducing dynamic and leakage power.
- Communication: Chip-to-chip links support rapid power-up and power-down, making communication energy proportional to transmitted bits.
- Numerical acceleration: Multiply-accumulate arrays were incorporated into the latest prototype to extend SpiNNaker2 beyond spiking-network simulation.
2. Brain simulation and other applications
SpiNNaker2 is intended to support larger and more complex spiking models while spanning multiple abstraction levels. Its throughput, low latency, and energy efficiency also target real-time mobile human-machine interaction.
- Brain simulation: SpiNNaker2 is expected to simulate significantly larger and more complex spiking neural networks than SpiNNaker1.
- Brain simulation: Numerical accelerators can support synaptic sampling and simulation of entire networks as mesoscopic black-box modules.
- Brain simulation: BioSpaun-like models could run from multicompartment neurons through spiking point neurons and rate-based or mesoscopic models.
- Other applications: SpiNNaker2 combines high-throughput machine learning, millisecond-latency sensor-actuator processing, and IoT-device-level energy efficiency.
- Other applications: These capabilities target real-time, mobile human-machine interaction, including tactile internet applications.