HWBRIEF
TU Dresden and Manchester researchers publish peer reviewed low precision (8 bit integer) measurements of SpiNNaker2, a 152 core Arm neuromorphic chip built to run both brain inspired spiking and conventional deep learning workloads on the same die.
Researchers at TU Dresden and the University of Manchester have published a peer-reviewed paper on SpiNNaker2, a many-core chip designed to run brain-inspired spiking neural networks and conventional deep-learning inference on the same die. The paper appears in the IEEE Open Journal of Circuits and Systems (IEEE Xplore; SemiEngineering digest).
SpiNNaker2 pairs 152 Arm Cortex-M4F cores on GlobalFoundries' 22FDX process with dynamic voltage and frequency scaling, LPDDR4 memory, and event-routed packets modeled on biological neuron firing. The authors, led by S. Scholze, describe the design as a way to bridge deep-network inference and event-based spiking workloads in one fabric — a claim the trade press had only gestured at since SpiNNcloud Systems' May 2024 commercial launch.
The headline numbers are INT8-specific: up to 4.5 TOPS in high-performance mode and up to 2.7 TOPS/W in high-efficiency mode. They validate the DNN side of the chip's hybrid claim. Spiking-mode validation is thinner; the architecture's promise depends on real workloads that combine both modes, not peak TOPS.
The commercial side is small. SpiNNcloud sells server boards with 48 SpiNNaker2 chips each, with early customers including Sandia National Laboratories. The chip is sold as a research platform, and the paper formalizes what it can do on conventional benchmarks. Whether event-based and conventional workloads actually coexist usefully on the same die is still open.