Google released more details about its AI supercomputer on Tuesday. Compared to Nvidia’s A100 chip, Google’s TPU architecture is said to be 1.7 times faster. It is also almost twice as efficient, Reuters reports.
Google is still using Nvidia’s technology, albeit on a smaller scale than before. Only 10% of Google’s AI training still takes place on chips that were not designed in-house. The supercomputer in question is used to process the datasets that chatbots, image generators and many other AI applications depend on. These calculations can take months to complete. Training Google’s most advanced large language model (PaLM), for example, took 50 days to complete.
Flexibility
A point against Google’s claim of superiority is that the company only compared its own TPUv4 technology to Nvidia’s older A100 chip. Nvidia already has a faster version on the market with the H100, a far stronger opponent to the TPUv4. These companies will doubtlessly continue to swap places in the coming years when it comes to state-of-the-art AI accelerators.
Google is particularly pleased with the flexibility the TPUs seem to have. For example, the architecture is able to work around broken components on the fly and features an adjustable topology, depending on what a machine learning model needs.
Tip: Nvidia’s new chips to power generative AI platform in Google Cloud
14 hours ago
