Google Unveils TPU V4 Supercomputer, Outshining Competitors

Yesterday, April 5, Google published detailed information about one of its new supercomputers with artificial intellect. The company claims that development bypassing speed and effectiveness competing NVIDIA systems.

While the “Green Company” dominates the market for training and deploying artificial intelligence models (more than 90% of the market), Google has been emphasized since 2016 since 2016 to develop its own high -performance chips, called tensor processors (Tensor Processing Unit, TPU) .

Nevertheless, over the past decade, Google has done a lot for the development of artificial intelligence. Employees of the American corporation have developed many key technologies that are currently used everywhere. However, from the point of view of the commercialization of its developments, according to many, the company is significantly lagging behind competitors.

In recent years, the company has not been the most favorable atmosphere inside the company, forcing developers and engineers to present such a product that would immediately prove to everyone that Google does not slow down and remains one of the leading companies in the field of artificial intelligence technologies.

Models and products with artificial intelligence, such as Google Bard or Openai Chatgpt, operating on NVIDIA A100 chips, require a large number of computers and thousands of chips used together to teach and work and work of II models. These computers are operated around the clock for many weeks or even months, which definitely imposes some inconvenience.

Even on Tuesday, Google said that it has created a single system with more than 4000 processors, designed for effective functioning, as well as the rapid training of artificial intelligence models. Google supercomputer based on TPU, called TPU V4, “1.2-1.7 times faster and consumes 1.3-1.9 times less energy than NVIDIA A100. The performance, scalability and accessibility is made by TPU V4 supercomputers with working horses For use in conjunction with large language models, ”says Google researchers.

However, the results of the tests of the new Google chips were not compared with the latest H100 chip from NVIDIA. It is more modern and most likely, the chips from Google on all fronts will cost. According to NVIDIA, H100 provide at least 4 times greater performance than the previous generation. Nevertheless, when NVIDIA has just demonstrated a new chip, Google has already introduced a ready -made supercomputer. So far, Google is ahead, but the situation may fundamentally change if the “green” in the near future present their own supercomputer.

/Reports, release notes, official announcements.