Google says its fourth-generation tensor processing unit (TPU) for training AI models is faster and more power-efficient than Nvidia's A100 chip.
In a new paper,
Google researchers detailed the company's latest and most advanced
TPU-based supercomputer, which has been operating since 2020.
- The AI system was used to train Google's Pathways Language Model (PaLM), a competitor to OpenAI's GPT series.
- Google
researchers made the claim that the "performance, scalability, and
availability " of TPU v4 supercomputers make them "the workhorses of
large language models."
- According to researchers, the TPU v4
supercomputer comprises 4,096 TPUs and other custom components. It's
~10x faster overall than the previous version.
- In the paper,
they wrote that for similarly sized computing systems, TPU v4 "is
1.2x–1.7x faster and uses 1.3x–1.9x less power than the Nvidia A100,"
Nvidia's data center GPU for machine learning tasks.
- Google did not compare TPU v4 to Nvidia's H100, the successor to the A100, saying it was released later and created using more advanced technology.
