Google has unveiled new information regarding the supercomputers it uses to train its artificial intelligence models. According to the tech giant, its custom-designed Tensor Processing Unit (TPU) chips are faster and more energy-efficient than comparable systems from Nvidia.
These chips are utilized in over 90% of Google’s work on artificial intelligence training, powering tasks such as generating images and responding to queries with human-like text. Google published a scientific paper detailing how it combined over 4,000 of these chips into a supercomputer using its custom-built optical switches to connect individual machines.
The connections are essential for training large language models that have grown significantly in size and cannot be stored on a single chip.
By allowing for easy reconfiguration of connections between chips, Google said its supercomputers can avoid problems and improve performance. The tech giant also revealed that its chips are up to 1.7 times faster and 1.9 times more power-efficient than comparable systems based on Nvidia’s A100 chip. Google suggested that it is working on a new TPU that could compete with Nvidia’s H100, but did not provide any further details.