We aim to develop specialized analog hardware bringing a new solution for performing training and inference of AI systems.
Embryo S.L.A.P proje... - Project
Analog computing came before digital, but now virtually everything is done digitally. In the last decade, much computing power has been allocated to training massive AI systems, performing on the order of 10^25 operations per training phase. These operations are performed on digital GPUs, located in superclusters of up to 200,000 GPUs. Needless to say, this has had large and unwanted consequences for power usage, and has locked out small companies from entering the AI market, given the massive compute disadvantage they are in.
Analog computing has been phased out due to its lack of precision and its limited flexibility. AI systems are inherently probabilistic, the lack of precision is therefore a non-issue. Moreover, the vast majority of operations in AI are multiplications and additions, meaning the lack of flexibility is not a significant problem either. Implementing AI with analog computing can result in significantly faster and more energy-efficient processes compared to digital systems.
Academic Supervisors