For the record this is alarmist bs. The article literally concedes that the power needs of AI haven’t been growing at a rapid rate (like crypto mining did).
Besides, advances in analog methods of computing specifically for the matrix multiplication that makes up the bulk of computing power needed are something like 15-60x more efficient while operating at a lower overall speed. When used together however, these analog chips can produce just as much computational power as digital chips at a fraction of the energy consumption. The chips are only really useful for training AI though, as they’re specifically engineered for matrix multiplication, so don’t expect analog computing to come back in other areas too
For the record this is alarmist bs. The article literally concedes that the power needs of AI haven’t been growing at a rapid rate (like crypto mining did).
Besides, advances in analog methods of computing specifically for the matrix multiplication that makes up the bulk of computing power needed are something like 15-60x more efficient while operating at a lower overall speed. When used together however, these analog chips can produce just as much computational power as digital chips at a fraction of the energy consumption. The chips are only really useful for training AI though, as they’re specifically engineered for matrix multiplication, so don’t expect analog computing to come back in other areas too