SpiNNcloud Systems says it is making commercially available a hybrid AI high performance computer system based on an architecture pioneered by Steve Furber, one of the designers of the original Arm processor.

SpiNNaker2 is based on the principles of the human brain, the company says, and uses a large number of low-power processors to more efficiently process AI and other workloads.

Specifically, the server board in a system carries 48 SpiNNaker2 chips, each of which has 152 Arm-based cores plus a variety of hardware accelerators, including distributed GPU-like units, all intended to boost neuromorphic, hybrid, and mainstream AI model processing.

The commercial system is designed to be scalable via a multi-rack setup, each composed of 90 SpiNNcloud server boards. According to the company, this configuration is capable of emulating at least 10 billion neurons. As a supercomputer also designed for machine learning workloads, it claims the system is capable of up to 0.3 exaops, where an exaop is 10¹⁸ operations per second.

SpiNNaker2 will be initially available as a cloud-hosted platform in the second half of 2024 to enable early customers to get access to the technology as cost efficiently as possible, SpiNNcloud says.

But it also reckons that full production systems will be ready to ship in the first half of 2025, and that institutions such as Sandia National Laboratories in the US, Technical University of München and Universität Göttingen in Germany, are already in line to receive theirs. No hint on pricing has been offered, but we have asked.

The SpiNNaker1 architecture was developed at the University of Manchester, as part of the €1 billion EU-funded Human Brain Project. Furber, who had a hand in the design of the BBC Micro and the Arm processor while at Acorn Computers, played a key part in this.

SpiNNaker received an €8 million (c $8.15 million) grant in 2019 to develop a second-gen chip based on the design.

The company behind the chip, Dresden, Germany-based SpiNNcloud, claims the SpiNNaker2 system has evolved from this predecessor to become an event-based platform for AI while "significantly enhancing all its original features."

The architecture has been developed to try to address the limitations that the developers see with the three main types of AI, comprising Deep Neural Networks, Symbolic AI/Expert Systems and Neuromorphic Models. Not least of these is the "ridiculous amount of energy" required to train current AI models, Furber states in a promo video for the launch of SpiNNaker2.

SpiNNcloud Systems co-founder and co-CEO Hector Gonzalez says the company's approach to artificial intelligence is via "brain-inspired supercomputer technology."

He says that a hardware platform enabling rule-based engines highly coupled with machinery to accelerate deep neural networks, plus one capable of operating efficiently like the human brains, would have the potential to revolutionize the large-scale implementation of AI models, and this is what SpiNNaker2 is intended to be.

In contrast to GPU-based solutions, SpiNNaker2 uses a large number of asynchronous low-power units to tackle workloads, leading to greater versatility, energy-proportional operation, and lower operational cost, he claims.

SpiNNaker2 sounds impressive, but the proof of the pudding is in the eating, and we await news of how users find it delivers for real applications, especially at those institutions investing in their own units.

SpiNNcloud will be at the ISC High Performance 2024 event in Hamburg next week, where attendees will be able to visit the company's booth. ®

https://www.theregister.com//2024/05/08/spinncloud_systems_unveils_armbased_neuromorphic/