Currently, data centers are estimated to account got 1-2% of global power consumption, and this number is expected to rise to 3-4% by the end of the decade. For a large part, this increase can be attributed to the recent developments in AI and its use of GPUs. Even though computations on GPUs are significantly more power efficient than on CPUs, the scale of computation needed to train and apply AI models is staggering and doubles every nine months.
Neuromorphic Computing is seen as an energy efficient alternative to GPU-based AI. This approach uses artificial neurons to do computations and is inspired by the structure and function of the human brain (which only requires about 0.3 kWh of energy per day; the equivalent of ~100 ChatGPT-4 queries).
There is currently significant investment in Neuromorphic Computing, both at fundamental level (such as materials and chip design) and applied level (algorithms, programming frameworks and applications). Several experiments chipsets exist (i.e., as Loihi 2 and SpiNNaker2) as well as large scale systems (i.e, spinncloud, Hala Point and Deep South). Many additional examples can be found on the Open Neuromorphic site. Within the Netherlands, initiatives like NL-ECO are exploring the benefits of this approach.
The goal of this project is to get an overview of the state-of-the-art in Neuromorphic Computing. What hardware is currently available? Is the software ecosystem mature enough? For which applications can Neuromorphic be applied. Can we get access to suitable testbeds to experiment with this paradigm? What opportunities do we see, and which downsides?
