Training artificial intelligence (AI) models not only requires high costs for purchasing equipment and building data centers, but also requires high electrical power. This is one of the reasons for the anti-AI movement because it is seen as a technology that destroys the natural environment. IBM today announced their research can reduce power usage by 5 times when training AI and also speed up the process by up to 77%.
The weakness in the existing system is the transfer of data between components that still use copper. Limiting the rate of electron transfer in copper forms a maximum speed limit that can be achieved. So IBM researchers developed co-packaged optics (CPO) technology that replaces copper connections with connections using optical polymer fibers.
Via the CPU, broadband data transfer at very fast rates can be done between the chip, circuit board and display. Because data is transferred using light, connections of hundreds of meters are still reasonable compared to using copper cables. Data transfer rates up to 80 times faster can be achieved using CPO
What's more important is that power usage can be reduced fivefold with IBM saying it equates to power usage for 5000 homes for a year. Quick transfer also managed to reduce the time to train an AI model from 3 months to 3 weeks.