Google recently announced the launch of its new Gemini artificial intelligence models, accompanying this event with the release of the latest version of its flagship tensor processor (TPU) for training and withdrawal of AI. This Google step is seen as an attempt to compete with the leading graphic processors (GPU) from NVIDIA.
The latest version of the TPU includes 8.960 chips for each node (part of the system), compared with 4.096 in version V4, and is four times more scalid in terms of accessibility of Flops to the node. New nodes provide throughput in 4.800GBPS and have 95GB high -speed memory (HBM) against 32GB HBM RAM in TPU V4.
The difference between the NVIDIA H100 and Google TPU V5P in speed: Google does not offer its TPUS for other companies for purchase, they are used exclusively inside the company for its own products and services. For a long time, Google TPUS was used to support services such as Gmail, YouTube and Android, and the latest version was also used to teach Gemini.
TPU V5P from Google 2.8 times faster teaches large language models than the TPU V4, and offers 2.1 times more values for money. Although the intermediate version of the TPU V5E, released earlier this year, offers the greatest value, it is only 1.9 times faster than TPU V4, which makes TPU V5P the most powerful option.
TPU V5P is even prayed enough to compete with the widely popular GPU H100 from NVIDIA, one of the best graphic cards for working with AI. This component processes working loads four times faster than the GPU A100 from NVIDIA, according to Data companies.
In the meantime, TPU V4 from Google, according to the study , published in April, 1.2-1.7 times faster than the A100. Preliminary calculations show that the TPU V5p is approximately 3.4-4.8 times faster than the A100, which puts it on a par with a higher or even higher than H100, although more details are needed for final conclusions.
.