Google disclosed new information related to the supercomputers it used to train its AI models on Tuesday, saying that they are faster and use less energy than comparable Nvidia Corporation systems.
The Tensor Processing Unit (TPU), a specialized processor, is employed for more than 90% of Google’s AI training. In order to make AI models perform better on tasks like producing human-sounding replies to questions or creating images, data is fed through the models during training.
Google (TPU) is now in its fourth generation. Google detailed how it connected over 4,000 TPUs to form a supercomputer in a scientific article released on Tuesday. The company has developed tailored optical switches to simplify the linking between different machines.
The research claims that Google’s supercomputer is up to 1.7 times faster and 1.9 times more energy-efficient than a system developed using a concurrently available fourth-generation TPU and Nvidia’s A100 CPU.
Google has asserted that it did not directly compare the two circuits since Nvidia’s most recent flagship H100 processor, which was unveiled after Google’s fourth-generation Tensor Processing Unit (TPU), is built using modern innovations.