Google has recently announced the release of its new artificial intelligence models known as Gemini, accompanying this with the release of the latest version of its flagship tensor processor (TPU) to teach and withdraw artificial intelligence (AI). Google views this step as a strategic attempt to compete with graphic processors, especially market leaders - graphic processors (GPU) from NVIDIA.
The TPU V5P, which is the most powerful Accelerator of AI from Google, was involved in working on AI Hypercomputer, a supercomputer architecture, specially designed to perform artificial intelligence tasks. This is different from ordinary supercomputers that are traditionally used for scientific calculations.
The latest version of TPU includes 8.960 chips per system of the system, which is a significant increase compared to 4.096 in the previous version of V4. It is also four times larger than the FLOPS plan on the knot. New nodes provide a striking bandwidth in 4.800GBPS and 95GB high -speed memory (HBM) compared to 32GB HBM RAM in TPU V4.
An important difference between NVidia H100 and Google TPU V5p is that Google does not offer its TPUS for sale to other companies; They are used exclusively within the company for their own products and services. Google TPUS has been used for a long time to support a variety of services such as Gmail, YouTube and Android. The new version was also used to teach the Gemini model.
Google V5p TPU teaches larger language models 2.8 times faster than TPU V4, and offers 2.1 times more value than money. Compared to the intermediate version of TPU V5E, released earlier this year, TPU V5P is the most powerful option, even if the V5E version offers the highest value that becomes 1.9 times faster than TPU V4.
Interestingly, the TPU V5P has proven to be powerful for competition with a widely used H100 graphic processor from NVIDIA, which is four times faster than the workloads than GPU A100 from NVIDIA, according to the company.
However, TPU V4 of Google, according to a study published in April, is 1.2-1.7 times faster than A100. Previous calculations indicate that TPU V5P is approximately 3.4-4.8 times faster than A100, placing it at or even higher than H100. However, final conclusions require more detailed tests.
