By Alessio Jacona*
Creating a hardware platform that allows the development of artificial intelligence that is much more powerful and efficient than the current one. This is the objective with which the giant Intel created and today presents Hala Point, the largest neuromorphic system in the world with 1.15 billion silicon neurons.
To put it very simply, a neuromorphic system is an experimental computer whose hardware as well as software structure is inspired by the functioning of the human brain, where instead of transistors we find silicon neurons, and which is much more efficient than traditional processors in making, for example, computers work. Large Language Model like ChatGPT. Indeed, according to Intel, machines of this kind will soon make it possible to better use artificial intelligence in the most diverse fields: from solving scientific and engineering problems to logistics to managing the infrastructure of smart cities and more
Hala Point is a large-scale system that improves on its predecessor Pohoiki Springs in every way, and uses Intel's new Loihi 2 processor with one million neurons. To give an idea, the previous version, simply called Loihi and presented in 2018, only had 131,000 neurons. The result is more than ten times greater neuronal capacity and up to 12 times greater performance in a system “that combines the efficiency of deep learning with new brain-inspired learning and optimization capabilities,” as Mike Davies, director of Neuromorphic, explains Intel Labs Computing Lab.
Let's run some numbers: the Hala Point neuromorphic system consists of 1,152 Loihi 2 processors in a six-rack-unit data center chassis that isn't even that big after all, given that it is comparable in size to that of a microwave oven . The system supports up to 1.15 billion neurons and 128 billion synapses spread across 140,544 neuromorphic processing cores that consume a maximum of 2,600 Watts of power. The system also includes more than 2,300 x86 processors for auxiliary computations.
The development of the system is headed by Intel Labs, which confirms that - although advanced - the project is still in the full research and development phase. Meanwhile, Intel tells us that its capacity is equal to 30 million billion (quadrillion) operations per second, or 30 petaops, with an efficiency that exceeds 15 trillion operations per second per Watt (TOPS/W) during running conventional deep neural networks. A result that on paper makes the system competitive if not even superior to GPU and CPU-based architectures.
As already mentioned, Hala Point is still a prototype intended for research: in the coming months it will be used by researchers at Sandia National Laboratories to work on solving scientific computing problems related to device physics, computer architecture and information technology.
«Conducting research with a system of this size - explained Craig Vineyard, Hala Point Team Lead at Sandia National Laboratories - will allow us to study calculation, modeling, simulation and data analysis capabilities to keep pace with evolution of artificial intelligence."
In a rapidly developing sector such as AI, the growth of deep learning models to trillions (trillion) parameters is calling into question sustainability in artificial intelligence, while highlighting the need to innovate at the lower levels of the hardware architecture. Neuromorphic computing is a fundamentally new approach that was created to solve this problem. It is based on insights from neuroscience, which integrate memory and computation with highly granular parallelism to minimize data movement. The result is that Loihi 2-based systems can perform AI inference (the process by which an AI model makes decisions based on new data) and solve optimization problems using 100 times less energy, at up to 50 times speed superior, compared to conventional CPU and GPU architectures.
*Journalist, innovation expert and curator of the ANSA.it Artificial Intelligence Observatory